DEV Community: Shamuddin

ZONOS3-Mini vs Fish Audio S2 Pro: Open TTS Verdict

Shamuddin — Sun, 15 Nov 2026 04:44:56 +0000

ZONOS3-Mini is the better pick if you need voice cloning on a phone, a laptop or a machine with no GPU, and Fish Audio S2 Pro is the one to check only if you are already committed to a server-side stack. Zyphra released ZONOS3-Mini on 9-10 November 2026 as a sparse Mixture of Experts model with 800 million total parameters, about 100 million of them active per token, under an Apache 2.0 licence that permits commercial use. The weights are on Hugging Face, and the model can run entirely inside a WebGPU-capable browser. Every quality number cited here is Zyphra's own.

TL;DR

ZONOS3-Mini is a real-time text-to-speech model with zero-shot voice cloning, released under Apache 2.0, with weights on Hugging Face.
It runs in the browser via WebGPU with no install and no server. Zyphra reports under 570 ms to first audio on a Pixel 9.
Four self-hosted runtimes exist: SGLang, MLX, C++/GGML and the browser build, all in the Zonos3 GitHub repository.
On Zyphra's own benchmark it beats Fish Audio S2 Pro and MOSS-TTS v1.5; nobody independent has reproduced this.
Reference-clip cloning scores well below the headline, so test with your own voices.

What is ZONOS3-Mini and why does the size matter?

According to the announcement page, Zyphra trained it on 112 NVIDIA H100 GPUs in 2.5 days - a modest budget for that feature list.

The size is the point. Its predecessor, ZONOS2, used 8 billion parameters and an 86 Hz audio codec. ZONOS3-Mini activates about 100 million parameters per token and uses a 15 Hz codec, so far less computation is needed for each second of speech. That reduction is what makes phone-class hardware viable.

How does ZONOS3-Mini work under the hood?

The GitHub README describes an 11-layer language model that mixes gated-DeltaNet layers with global attention, with one global layer in every four. Text is tokenised with SuperBPE. Audio is represented as dual-codec codes at 15 Hz and decoded by a causal multi-resolution iSTFT decoder that outputs 44.1 kHz audio.

About 27 built-in voices ship grouped by language. You can clone from a raw reference clip (encoded into the same codes) or a 192-dimension MECT-B2 speaker embedding, steer pace and emotion word by word, and text normalisation covers 9 languages including English, Mandarin, Japanese and Korean.

Where can ZONOS3-Mini run, and how fast is it?

Latency below is self-reported by Zyphra on its announcement page; runtimes are listed in the Zonos3 repository.

Runtime	Hardware	Install cost	Reported time to first audio
WebGPU in browser	Phone, laptop	None; browser downloads and caches the model	Under 570 ms on a Pixel 9; under 150 ms on an M3 MacBook Air
MLX	Apple Silicon	Python environment	Not stated separately by Zyphra
SGLang	NVIDIA server GPU	Server setup; tested on H100	80 ms on one H100
C++/GGML	CPU, CUDA, Metal, Vulkan	Build the server and CLI	Not stated separately by Zyphra

The browser route is unusual: the model downloads and caches once, then inference stays on the device, so no audio or text leaves the page. WebGPU support varies by browser and OS, so check yours first.

How do I try ZONOS3-Mini in three steps?

Three starting routes, in order of effort. Flags change between releases, so follow the repository README rather than copying commands from a blog.

Browser playground. Open the WebGPU build linked from the Zyphra page in a recent Chrome-class browser, wait for the one-time model download, then generate a short sentence to confirm latency on your device.
SGLang server. On an NVIDIA GPU, start the SGLang runtime from the repository. It exposes an OpenAI-compatible /v1/audio/speech endpoint, so existing client code should need little change beyond the base URL and voice name.
GGML on CPU. Build the C++ server or CLI with the CUDA, Metal or Vulkan backend if you have one - the route for headless machines where a browser is not an option.

If you want zero hosting, Zyphra Cloud serves it on NVIDIA H200s; I could not verify pricing, so check the page.

Is ZONOS3-Mini better than Fish Audio S2 Pro?

On Zyphra's numbers, yes. On its ZTTS2-Eval benchmark, Zyphra reports a capability index of 66.2 against 61.7 for Fish Audio S2 Pro and 60.2 for MOSS-TTS v1.5. On the clean set, word error rate is 7.79 for ZONOS3-Mini against 11.06 for Fish Audio S2 Pro, as summarised by Marktechpost.

Three cautions apply: the benchmark is vendor-built, no independent reproduction had appeared as of 11 November 2026, and the same index falls to 44.7 for reference-clip cloning. If cloning fidelity drives your decision, run your own comparison on your own recordings.

What is the verdict, and who should choose what?

For zero-GPU or phone-local voice cloning under a permissive licence, ZONOS3-Mini is the strongest open option I can point to today. The combination of Apache 2.0 terms and a phone-class WebGPU build is not something I found in the models Zyphra compares against, though I have not audited their licences line by line.

Why we picked it for the verdict: per our n=72 zero-difficulty keyword audit, measured with DataForSEO volume and difficulty data on November 9, 2026, found 20 keywords where no established competitor holds the result set. Model-name intent like this is exactly that class - the comparison is decided by hands-on facts, not by search authority. Choose ZONOS3-Mini for on-device synthesis, privacy by construction, or a self-hosted server without a large GPU. Stay on Fish Audio S2 Pro or MOSS-TTS v1.5 if they already meet your quality bar and you need no edge execution. Do not switch production on a vendor benchmark alone.

For other angles on the same decision, see our comparisons of Voicebox vs ElevenLabs, Breeze TTS 2 vs ElevenLabs licensing and VoxCPM2 vs ElevenLabs self-hosting; our free local dictation on Mac guide covers the input side.

What should I watch for next?

Two things will settle the claims: independent WER and capability reproductions on public test sets rather than ZTTS2-Eval, and any response from the Fish Audio or MOSS-TTS teams, since a release from either resets the comparison.

FAQ

Q: Is ZONOS3-Mini free for commercial use?

A: Yes. The weights ship under the Apache 2.0 licence, which permits commercial use. Read the licence text in the repository before shipping.

Q: Does ZONOS3-Mini really run on a phone?

A: Zyphra reports under 570 ms to first audio on a Pixel 9 via WebGPU. That is self-reported; real results depend on your browser, device and the first-load download.

Q: Do I need a GPU to run ZONOS3-Mini?

A: No. The browser build uses WebGPU, and the C++/GGML runtime can run on a CPU. Only the SGLang server path expects an NVIDIA server GPU, and it was tested on an H100, per the Zonos3 repository.

Q: How many parameters does ZONOS3-Mini have?

A: 800 million total, about 100 million active per token, because it is a sparse Mixture of Experts, per the Zyphra announcement page.

Q: Can ZONOS3-Mini clone any voice from a short clip?

A: It clones zero-shot from a reference clip or speaker embedding. Zyphra's own index falls to 44.7 for reference-clip cloning, so quality varies; only clone voices you have permission to use.

Q: Which languages does ZONOS3-Mini support?

A: Text normalisation covers 9 languages including English, Mandarin, Japanese and Korean, per the README; built-in voices are grouped by language.

Neural Network Physics Simulation: WorldParticle vs Solvers

Shamuddin — Sun, 15 Nov 2026 04:44:36 +0000

By Sham Uddin. Last verified: November 10, 2026.

On November 9, 2026, MIT News described WorldParticle, a neural network physics simulation that models sand, cloth, fluids, elastic solids and proteins with one architecture instead of a separate hand-coded solver for each. Our verdict: for exploratory work where speed matters, such as sampling protein motion or prototyping soft-robot contact, WorldParticle beats conventional numerical simulators, which MIT reports it runs about 100 times faster on protein dynamics. For engineering sign-off where every result must be validated against known physics, the established solvers still win until independent replication exists. The primary source is the MIT News release.

TL;DR

WorldParticle is one transformer-based model that simulates many materials, from sand to protein molecules, by tracking particles.
MIT says it modelled protein structure changes about 100 times faster than numerical simulators, and it was more accurate than the numerical methods it was compared against.
The paper is a preprint on arXiv (2605.15305) and will be presented at SIGGRAPH Asia.
It generalises to unseen materials, forces and boundary conditions, according to the authors.
It does not replace validated solvers today. Treat it as a fast first-pass tool, not a certified one.
No product or pricing exists. This is open-access academic research.

What is WorldParticle and who built it?

WorldParticle is a research system from MIT, the University of British Columbia (UBC) and collaborators at Georgia Tech, Inria, Meta, UPenn, the University of Utah, UCLA and the University of Hong Kong. The co-lead authors are Minghao Guo (MIT postdoc), Caoliwen Wang and Siyuan Chen (UBC graduate students). Peter Yichen Chen (UBC) is co-corresponding author, and Wojciech Matusik leads the Computational Design and Fabrication Group at MIT CSAIL.

Today, scientists pick a different solver for each physical process: one for flowing water, another for protein folding, another for cloth. WorldParticle replaces that patchwork with a single learned model. As Chen put it in the MIT release, the aim is "a GPT moment for simulation".

How does neural network physics simulation work in WorldParticle?

The model represents every material as thousands to millions of small 3D particles, a Lagrangian representation, and predicts where each particle moves next. MIT describes the output in one-second increments. Chen compares it to a chat model, except that it predicts how particles move one step at a time rather than one word at a time.

According to the arXiv preprint, the design is prediction-correction:

An explicit predictor advances particles under known external forces, such as gravity.
A learned corrector estimates the residual position and velocity updates the predictor misses.
Inside the corrector, a particle tokenizer turns particles into tokens, and a super-token encoder merges them hierarchically into compact super tokens, halving the token count at each level.
A super-token decoder lifts the result back to full particle resolution using cross-attention.

The merging step matters for cost. Attention over millions of particles is expensive, and working on a compressed set of super tokens keeps the learned part tractable. Because the model learns dynamics from data instead of encoding equations, it can adapt to materials and environments absent from its training set.

How does WorldParticle compare with numerical solvers?

Aspect	WorldParticle	Conventional numerical solvers
Scope	One architecture across six dynamics categories	One solver per physical process
Basis	Learned from data, with an explicit force predictor	Hand-coded equations
Protein dynamics speed	About 100x faster, per MIT	Baseline
Validation record	Preprint, not yet independently replicated	Decades of validation
Best use	Fast exploration, design iteration	Certified engineering results

The six categories are cloth, elastic solids, Newtonian fluids, non-Newtonian fluids, granular materials and molecular dynamics. In one granular demonstration, baseline methods left sand particles floating in mid-air, while WorldParticle captured the collapse into a pile.

Why is the 100x protein speed-up plausible?

The gain comes largely from step size. Yahoo Tech's reading of the paper reports that protein runs used 50-femtosecond steps, which is 100 times longer than the 0.5-femtosecond steps of conventional molecular simulation. Fewer steps to cover the same elapsed time means less compute.

The tradeoff is that a learned model taking large steps must be right about the physics it skips. Numerical solvers use small steps for numerical stability, and their errors are well understood. A neural model's errors are harder to bound, which is why independent benchmarks matter before anyone relies on it for drug-related or safety-critical modelling. Nothing in the source material supports medical use, and this article does not suggest it.

What could engineers and researchers do with it?

MIT points to two kinds of problem that normally need several simulators stitched together. One is a fluid sloshing against a deformable solid, such as oil hitting an elastic valve in a storage tank. The other is how sand or soil moves when a soft robot touches it. The paper also demonstrates interactive control, inverse design and learning from real-world manipulation data.

Longer-term applications proposed by the authors include RNA evolution, astrophysical phenomena such as merging galaxies, and a general-purpose simulation pipeline for materials and medicine discovery. Those are goals, not results.

Where does it sit among other AI science models?

WorldParticle differs from narrower scientific models. Some target one domain, such as protein structure or chemistry experiment planning. Our coverage of Hydrogen-1.0 versus GPT-6 Astra on science benchmarks and the BioBasicsBench wet-lab benchmark shows how those systems are tested on knowledge and planning. WorldParticle instead targets physical dynamics directly. This is also why we track who actually occupies these result pages: "We captured live top-10 organic results for 7 target queries to see who actually occupies the page rather than who claims to." (n=7, captured November 10, 2026).

For a biology-side view of how generative models are being applied and where the risks lie, see our piece on AI-designed viruses from Evo 2. On the graphics side, particle and point representations also appear in SplatAnything and 3D Gaussian splatting. For a look at how head-to-head model evaluations are built, read the GDPX benchmark comparison.

What are the limits and open questions?

It is a preprint (arXiv 2605.15305, first submitted May 14, 2026). The November 9 news is the MIT release and the SIGGRAPH Asia presentation, not a peer-reviewed journal result.
The accuracy claim compares against the numerical methods the authors chose. Independent tests on other problems are needed.
Generalisation to unseen materials is reported by the authors. Failure modes on very different physics are not yet documented in public benchmarks.
No error bounds are offered of the kind a certified solver provides.

The paper and project page are open access, so outside groups can test these claims: project page.

FAQ

Q: What is WorldParticle?
A: WorldParticle is a single deep-learning system from MIT, UBC and collaborators that simulates many materials, including sand, cloth, fluids, elastic solids and proteins, by predicting how thousands to millions of particles move step by step.

Q: Is WorldParticle faster than traditional simulation?
A: According to MIT postdoc Minghao Guo, it modelled protein structure changes about 100 times faster than numerical simulators. That figure comes from the authors and has not been independently replicated.

Q: Is WorldParticle peer reviewed?
A: The paper is an arXiv preprint (2605.15305) and will be presented at SIGGRAPH Asia. Treat the results as promising but not yet independently confirmed.

Q: Can I use WorldParticle today?
A: It is open-access research with a public project page and paper. There is no commercial product or pricing, so adoption would mean working from the published research.

Q: Will it replace numerical solvers?
A: Not soon. It suits fast exploration, but validated numerical solvers remain the standard where results need documented accuracy and error bounds.

Q: What materials can it simulate?
A: The paper covers cloth, elastic solids, Newtonian and non-Newtonian fluids, granular materials and molecular or protein dynamics, and reports generalisation to unseen materials within those categories.

This article was produced with AI assistance and checked against the primary sources linked above. Corrections log: none as of November 10, 2026.

Shadcn Improve vs Manual Subagents After 10,423 Stars

Shamuddin — Sun, 15 Nov 2026 04:44:15 +0000

Shadcn improve is the better choice than a hand-rolled Claude Code subagent split for most teams, because it turns the "expensive model plans, cheap model executes" pattern into a fixed contract instead of a habit you have to maintain. It reached 10,423 GitHub stars and 427 forks on 11 November 2026, with 333 stars gained in a single day. The manual route still wins for small, one-off jobs where installing anything is more effort than the task.

TL;DR

Shadcn improve is an Agent Skill that audits a codebase and writes implementation plans. It never edits source code; its only writes go to a plans/ folder (repo README).
It hit 10,423 stars at the 13:00 UTC scan on 11 November 2026 (Early Signal).
Verdict: adopt improve if you run agents on a codebase more than occasionally. Stay manual for a single throwaway task.
Install with npx skills add shadcn/improve, then start with /improve quick.
Last verified: 11 November 2026.

Why is shadcn improve trending again?

The repository was created on 10 June 2026 and passed 1.4K stars in its first 48 hours, according to a third-party review. That same review recorded 7,534 stars and 314 forks on 9 July. Today's numbers are well above that baseline.

Date	Stars	Source
9 July 2026	7,534	dev.to review
11 Nov 2026, 13:00 UTC	10,423	Early Signal
11 Nov 2026, 14:05 UTC	10,488	GitHub API read

Our analysis of the two recorded endpoints (Early Signal, dev.to review) puts the July-to-November growth at 2,889 stars, an average of about 160 stars a week, which makes the 333-star day roughly twice the repo's usual weekly pace. The 14:05 UTC row is our own check of the live repository state, a single API snapshot (n=1). Stars measure attention, not quality, so treat this as a reason to look, not a reason to adopt.

What does shadcn improve actually do?

Shadcn improve, written by the creator of shadcn/ui, splits AI-assisted maintenance into two jobs. The most capable model does the thinking: understanding the code, judging what matters, writing the specification. A cheaper model does the typing. In the project's own framing, the plan is the product.

The audit runs nine categories in parallel subagents: correctness, security, performance, test coverage, tech debt, dependencies and migrations, developer experience, docs, and direction. Every finding carries file and line evidence, so you can check a claim before approving any work.

What makes an improve plan different from a prompt?

Each plan is built to survive being handed to an agent that has never seen your repository (repo README):

Self-contained: all required context is inlined in the plan.
Verification gates: every step ends with a command and its expected output.
Hard boundaries: explicit STOP conditions tell the executor when to halt rather than improvise.
Version-stamped: each plan records the git commit it was written against, so stale plans are detectable.

Plans are plain markdown, so a human or any other agent can pick one up. That portability matters if your team also uses other tools, such as the setup covered in our GitHub Copilot in Xcode 27.3 guide.

How do I set up shadcn improve?

Run npx skills add shadcn/improve in any host that supports the Agent Skills format.
Open your agent in the repository and run /improve quick first. This is a short pass over the highest-confidence findings, and the review puts it at about five minutes. Standard mode takes about fifteen.
Reply with the findings you want planned, for example "plan 1, 3 and 5". Plans land in plans/ with an index.
Run /improve execute 001 to dispatch a cheaper executor into an isolated git worktree. The reviewer returns approve, revise (up to twice) or block. Merging stays with you.
Run /improve reconcile next session to refresh the backlog, and /improve branch before a pull request to audit only what the branch changed. Adding --issues publishes plans as GitHub issues.

How does it compare with a manual Claude Code subagent split?

You can approximate the same pattern by hand. Write a brief, have a strong model draft a plan, then start a cheaper subagent to execute it. Our piece on Claude Code subagents and cheaper status updates covers the cost logic of delegating routine work, and the thinking behind plan-first work is explored in verify-first versus build-first engineering.

Factor	Shadcn improve	Manual subagent split
Setup	One install command	None, but you write the brief every time
Plan consistency	Enforced by the skill's contract	Depends on whoever writes the prompt
Evidence	file:line on every finding	Whatever you ask for
Isolation	Git worktree per execution	Your responsibility
Flexibility	Bounded by its audit categories	Unlimited
Best for	Ongoing codebase upkeep	One-off or unusual tasks

The manual approach is not worse in principle. It is worse in practice once the third person on your team writes their own version of the brief. A fixed contract with STOP conditions and verification gates removes that variance.

What are the limitations?

The review is candid about the costs (dev.to). Deep mode on a 200K-line monorepo is still not cheap. Audits are limited to the playbook's categories, so unusual concerns can slip through. Very weak local models as executors need more revision rounds. Reconcile also assumes you return to the project regularly.

There is a process objection too: teams that live in Linear or Jira may not want plans as markdown in git. The --issues option partly addresses that. The SKILL.md file is at v1.0.0, and the repository has no formal GitHub releases, so pin the commit you install if you need reproducibility.

If you are choosing a wider platform for agent-driven coding, our open-source AI agent platform comparison is a useful companion.

Who should adopt shadcn improve?

Adopt it if you maintain a codebase over months, share agent workflows across more than one developer, or want auditable evidence behind each change. Stay manual if you only need a single refactor, or if your project has requirements the nine audit categories do not cover. Either way, start with /improve quick on a non-critical repository and judge the findings before trusting the executor.

FAQ

Q: Does shadcn improve change my source code?
A: No. The skill only writes to the plans/ folder. Code changes happen only if you run /improve execute, which uses an executor in an isolated git worktree, and merging remains your decision.

Q: How many stars does shadcn improve have?
A: It had 10,423 stars and 427 forks at the 13:00 UTC scan on 11 November 2026 (Early Signal), and a live API read at 14:05 UTC showed 10,488.

Q: Which agents can run shadcn improve?
A: Any agent that supports the Agent Skills format. Install it with npx skills add shadcn/improve. Because plans are plain markdown, other agents and humans can also use them.

Q: What is the difference between quick and standard mode?
A: Quick mode returns the top high-confidence findings and takes about five minutes. Standard mode produces the full findings table and takes about fifteen, per the dev.to review.

Q: Is shadcn improve free to use?
A: The project is MIT licensed (GitHub). You still pay for the model usage of whichever agent runs the audit and the executor.

Last verified: 11 November 2026. This article was researched and drafted with AI assistance and reviewed against the primary sources linked above. Corrections are logged on this page when we find them.

Anthropic IPO for Indian Investors: $250,000 LRS Limit

Shamuddin — Sun, 15 Nov 2026 04:43:55 +0000

The Anthropic IPO now has a price: on November 10, 2026, Anthropic, PBC set a preliminary range of $70 to $78 per Class A share for 825 million shares, raising up to $64.35 billion at the top end, according to its Amendment No. 2 to Form S-1 on SEC EDGAR. For Indian residents the practical answer is yes, you can buy listed foreign shares, but only through the Reserve Bank of India's Liberalised Remittance Scheme, which caps remittances at USD 250,000 per financial year; whether you get an allocation at the IPO price depends on your broker, not on RBI rules.

TL;DR

Range: $70.00 to $78.00 per Class A share, Nasdaq Global Select Market, ticker ANT (SEC filing).
Size: 825 million shares, so $57.75 billion at $70 and $64.35 billion at $78 (arithmetic on filed figures; top end also reported by Yahoo Finance).
Valuation: about $917 billion to $1.02 trillion on shares outstanding, up to $1.405 trillion fully diluted (Yahoo Finance).
Timing: pricing November 18 and first trading November 19 are reported, not filed (shattered.io IPO tracker).
India: the route is LRS, with tax collected at source of 20 percent above Rs 10 lakh for non-education, non-medical remittances in FY 2026-27.
Builders: nothing changes for the Claude API, models or pricing on the day.

What exactly did Anthropic file?

The offering is 825,000,000 Class A shares: 801,351,354 sold by Anthropic and 23,648,646 by existing stockholders, and Anthropic receives no proceeds from the stockholder portion. Underwriters hold a 30-day option on up to 123,750,000 additional shares, with Goldman Sachs & Co. LLC and Morgan Stanley & Co. LLC as representatives. The roadshow began the same day and the preliminary prospectus is public.

Two details matter beyond the headline. Seven technology companies have agreed to buy $26 billion of Class N non-voting shares at the IPO price, outside the public offering, so that money does not count toward the $64.35 billion. And the amendment reports run-rate revenue of about $81 billion as of October 31, 2026, up $10 billion from September 30. The quarterly figures from the earlier amendment are in our November 3 news digest, so I will not repeat them here.

How much is Anthropic worth at $70, $74 and $78?

The filing indicates about 13.09 billion shares outstanding after the sale, which StockTitan used to derive the valuation range. Our analysis of the filed share counts extends the arithmetic into the table below, dividing market value against the filing's $81 billion run-rate revenue.

Price per share	Gross offering (825M shares)	To Anthropic (801.35M shares)	Market value (13.09B shares)	Value per $1 of run-rate revenue
$70	$57.75 billion	$56.09 billion	about $917 billion	11.3x
$74 (midpoint)	$61.05 billion	$59.30 billion	about $968 billion	12.0x
$78	$64.35 billion	$62.51 billion	about $1.02 trillion	12.6x

The share counts come from the filing and the roadshow announcement coverage; the multiplication is mine. A price of about $76.50 crosses $1 trillion on outstanding shares, per StockTitan.

The fully diluted figure is higher because stock options, unvested restricted stock units and plan reserves count too: up to $1.405 trillion, including options on about 1.77 billion shares at a $2.91 average strike, as reconciled step by step by shattered.io. When you compare Anthropic with a listed peer, use the same share basis for both, or the comparison is meaningless.

Can Indian residents buy the Anthropic IPO?

Yes, within limits. Under the RBI Master Direction on the Liberalised Remittance Scheme, resident individuals may remit up to USD 250,000 per financial year, April to March, for permitted capital account transactions, which include buying listed foreign shares. That ceiling is about Rs 2.26 crore at the conversion used for this article.

What the scheme does not do is guarantee an allocation. Brokerages report a retail allocation target of about 20 percent, but retail IPO shares are distributed by the underwriters' selling groups, and most India-based buyers will get ANT only on the open market after trading starts, at whatever price it finds. Ask your broker which of the two it supports before you move money.

How do I invest in the Anthropic IPO from India under LRS?

Confirm access: check that your broker or international platform lists Nasdaq stocks for Indian residents, and whether it offers IPO allocations or only secondary-market buying.
Open the route: complete the LRS declaration with your authorised dealer bank or platform partner, stating the purpose as investment in listed overseas securities.
Check the tax: TCS is nil up to Rs 10 lakh and 20 percent above it for purposes other than education and medical in FY 2026-27. It is collected up front and can usually be claimed against your income tax liability; confirm with a chartered accountant.
Place the order: a limit order is safer than a market order on listing day, because the opening price can differ sharply from the offer price.

For scale: at the $78 top of the range, the full USD 250,000 LRS annual limit is about 3,200 shares, worth $24.9 million at that price, so the ceiling binds only for the largest family offices. This is an illustration of the rule, not tax advice.

Is the Anthropic IPO priced to rise on day one?

Nobody knows, and the range can change before pricing. Two reference points. First, the implied valuation of about $968 billion at the $74 midpoint (StockTitan arithmetic on the filing) sits just above the $965 billion investors put on Anthropic in May 2026 (Yahoo Finance), so bankers have pitched the IPO flat against that round rather than above it. Second, at $78 the market value of about $1.02 trillion sits roughly 12.6 times the $81 billion run-rate revenue, the same division shown in the table above - 12.0x at the $74 midpoint and 11.3x at $70. Run-rate is an annualised measure of recent revenue, not recognised revenue, and the company's losses are a separate matter.

What does the Anthropic IPO change for people building on Claude?

Nothing today. The roadshow announcement says nothing about products, and the Claude API, models and pricing are unchanged by a listing. If you are choosing a model, the decision still rests on cost and capability, as covered in our comparison of Claude Sonnet pricing against GLM 5.3 and our look at GPT-6 Cyber against Claude Opus 5.5. Indian users weighing subscriptions can see Claude Pro against Google AI Pro in India, and the company's partnerships in the country are in our piece on the Anthropic licensing deal with ANI.

Verdict for builders: do not change your stack because of the listing. What to watch over the coming quarters is how quarterly public-market reporting affects compute commitments and margins, particularly while open-weight models keep pressure on per-token prices. Keep a second provider integrated either way.

FAQ

Q: What is the Anthropic IPO price range?
A: $70.00 to $78.00 per Class A share, listed on the Nasdaq Global Select Market under ANT, per the SEC filing.

Q: How much will Anthropic raise?
A: Up to $64.35 billion gross at $78, or $57.75 billion at $70. Anthropic keeps the proceeds from 801,351,354 of the 825 million shares.

Q: When will Anthropic start trading?
A: Pricing on November 18 and first trading on November 19, 2026 are reported by CNBC-sourced coverage citing people familiar; the dates are not in the filing.

Q: How much can an Indian resident invest abroad?
A: Up to USD 250,000 per financial year under LRS, with TCS of 20 percent above Rs 10 lakh for most purposes.

Q: Does the IPO change Claude API pricing?
A: No. The roadshow announcement does not mention products or pricing.

Q: Is the $1 trillion valuation certain?
A: No. About $76.50 per share crosses it on shares outstanding, and the final price can land anywhere in or outside the range.

This article is informational and not investment or tax advice. Figures reflect filings and reports available on November 10, 2026. Corrections log: none to date.

Best AI Transcription Software Compared: Otter Wins

Shamuddin — Sun, 15 Nov 2026 04:43:35 +0000

Verdict: For most small teams, the best transcription software is Otter. Its free plan gives 300 minutes a month, and its paid entry tier costs $8.33 per user per month billed annually or $16.99 monthly, which is the lowest full commercial plan in this comparison. Choose Rev when a transcript must be verified by a human, Happy Scribe when language coverage or EU data residency decides, TurboScribe for high-volume file uploads, Sonix for irregular months, and Descript only when the transcript is how you edit audio or video. All prices below were checked on each vendor's pricing page on 10 November 2026.

TL;DR

Otter is the best default for meeting-heavy teams: cheapest paid entry, a free tier usable for real work.
Rev when accuracy must be guaranteed: human transcription from $1.99 per minute, 99%+ guaranteed.
Happy Scribe covers 150+ AI languages and holds data in Dublin.
TurboScribe is cheapest for bulk uploads; wasteful at a few hours a month.
At 10 hours a month on monthly billing, list prices run about $1.70 to $10 per hour - the pricing model matters more than the headline.

How do the best transcription tools compare at a glance?

	Otter	Rev	Happy Scribe	Sonix	TurboScribe	Descript
Free tier	300 min/mo, 30-min conversations	45 AI min/mo, English only	10 min for uploaded files	None	3 files/day, 30 min each	1 media hour/mo
Entry paid price	$16.99/mo, or $8.33 annual	$29.99/seat/mo, or $25.49 annual	$17/mo, or $8.50 annual	$10/hour pay-as-you-go	$20/mo, or $10 billed yearly	$24/mo, or $16 annual
Meter	Recording and import minutes	AI minutes	Uploaded minutes	Hours	Unlimited	Media hours
Languages	6	English core	150+	54+	98+	25
Human check	No	From $1.99/min	From $2.00/min	No	No	No
Best for	Meetings	Verified records	Languages, EU data	Sparse months	Bulk files	Editing by transcript
Key limitation	Few languages	Heavy legal/investigative positioning	Small allowances on entry plans	Costly at volume	No meeting bot or summaries	Counts raw recording time

Sources: Otter, Rev, Happy Scribe, Sonix, TurboScribe, Descript.

What does an hour of transcription actually cost?

Sticker prices hide how differently these tools count usage. The arithmetic below assumes a realistic month: 10 hours of audio, monthly billing.

Otter Pro: $16.99 divided by 10 hours is about $1.70 per hour. Imports are limited to 10 a month with meetings up to 90 minutes, so ten one-hour files fit, but longer recordings need splitting.
Descript Hobbyist: $24 for 10 media hours is $2.40 per hour. Annual billing at $16 brings that to $1.60.
TurboScribe Unlimited: $20 monthly is $2.00 per hour at this volume, and falls as you upload more (pricing).
Happy Scribe Pro: $29 for 600 uploaded minutes is $2.90 per hour, or $1.90 on the $19 annual rate.
Rev Essentials: $29.99 monthly is about $3.00 per hour, or $2.55 on the $25.49 annual seat. Those are AI minutes; human transcription is billed separately.
Sonix: pay-as-you-go is $10 per hour with no commitment. A steady 10 hours is better served by the $50 Advanced plan, which works out at $5.00 per hour at this volume.

The spread runs from roughly $1.70 to $10 per hour for identical work - Otter's worked rate against Sonix's pay-as-you-go (Otter; Sonix) - a sixfold gap. The cheapest option in this month is the most expensive in a month with no audio, which is why the pricing model deserves as much attention as the price.

Why is Otter the best transcription software for most small teams?

Otter combines the lowest paid entry price with a free tier that works for real use. The Basic plan is free with 300 minutes a month, 30-minute conversations and three lifetime imports, presented on the plan table without trial-style restrictions.

Pricing (checked 10 November 2026): Pro is $16.99 per user monthly or $8.33 annually. Business is $30 monthly or $19.99 annually, with unlimited meetings, meetings up to four hours and three concurrent meetings. Enterprise adds SSO, SCIM and an HIPAA add-on, priced on request.

Strengths: live transcription, speaker identification, team vocabulary, and mp3/txt/pdf/docx/srt exports; Pro connects to Salesforce, HubSpot and Zapier.

Limitations: six languages only (English, Spanish, French, German, Japanese, Chinese). Free minutes do not roll over, and there is no human verification for poor audio or heavy accents.

When is Rev worth paying for?

Rev is worth it when the cost of an error exceeds the cost of a human check. Its human transcription starts at $1.99 per minute, 99%+ accuracy guaranteed, delivered in 12 hours or less - so one hour of human work costs about $119.40 (Rev), and the AI tier is stated at 95%+ on the same page.

Pricing: the free plan gives 45 English-only AI minutes a month. Essentials is $29.99 per seat monthly or $25.49 annually with 5,000 AI minutes and up to three users; Pro is $59.99 monthly or $47.99 annually with 10,000. Legal formatted transcripts are per page, not per minute: $2.25 rough, $2.50 premium.

Compliance: Rev lists CJIS, SOC 2 Type II and HIPAA.

Limitations: subscriptions cover AI minutes and human work bills separately; Rev's positioning now leans legal and investigative, so plans feel less tailored to office meetings.

Is Happy Scribe the best choice for languages and EU data?

Yes, on both counts. Happy Scribe supports 150+ AI languages and dialects, the most here, and holds data in Dublin for EU residency (independent comparison).

Pricing: the free plan allows unlimited Notetaker recordings of up to 45 minutes and just 10 free minutes for uploaded files. Basic is $17 monthly or $8.50 annually for 120 uploaded minutes, with overage at $0.20 per minute. Pro is $29 monthly or $19 annually for 600 minutes and three seats. Business is $89 monthly or $59 annually for 6,000 minutes and five seats. Human proofreading starts at $2.00 per minute.

Limitations: the 120-minute entry allowance runs out quickly, and the free upload allowance is closer to a trial than a working tier.

Which option suits bulk uploads: TurboScribe or Sonix?

TurboScribe suits steady, high-volume uploads. Unlimited costs $20 monthly or $10 a month billed yearly, takes files up to 10 hours or 5 GB, 50 at a time, in 98+ languages on Whisper models; the free plan allows three files a day up to 30 minutes. There is no meeting bot or AI summaries - it is upload-and-export.

Sonix suits uneven demand. Pay-as-you-go is $10 per hour with no subscription; Core is $25 monthly for 5 hours, Advanced $50 for 20 hours, Pro $80 for 40 hours, overage $10 per hour on any plan. It covers 54+ languages with a custom dictionary and a medical model, but has no free plan and gets expensive at volume next to TurboScribe.

Is Descript a transcription tool or an editor?

Descript is an editor first: its transcript is the editing surface, so deleting words cuts the recording. Hobbyist is $24 monthly or $16 annually for 10 media hours, Creator $35 or $24 for 30, Business $65 or $50 for 40; the free plan gives one media hour a month and watermarked 720p export.

The catch is the meter: Descript counts raw recording time, not finished length - a two-hour session that becomes a 40-minute episode still uses two hours. If you only need text, you are paying for editing features you will not use.

How should you choose? A short decision path

We weighted real cost at a typical month's volume, pricing-model fit, language coverage, accuracy path and free-tier usability, and published no numeric scores because the weights are a judgement call. The rankings, criterion by criterion:

Cost at volume: Otter, then Descript and TurboScribe, then Happy Scribe and Rev, Sonix pay-as-you-go last.
Pricing-model fit: TurboScribe for steady volume, Sonix for sporadic months, Otter for meeting-driven work.
Languages: Happy Scribe, TurboScribe, Sonix, Descript, Otter, Rev Essentials.
Accuracy path: only Rev and Happy Scribe sell human verification.
Free tier: Otter first; Sonix offers none.

Otter leads or sits near the top on three of five and has no outright failure on the others, which is why it wins. Start from the question "what do I record?" and switch only if one of the exceptions above applies. For tools that also write summaries and action items, see our comparison of AI meeting assistants for small businesses.

What are the credible alternatives?

Self-hosted Whisper: free software, but the hardware and maintenance are yours.
Meeting-assistant suites: add summaries and CRM sync; see our small-business meeting assistant comparison.
Support-call analysis: for customer calls, our AI customer service software guide covers tools that transcribe within a wider workflow.
Speech generation (the opposite direction): we compared Voicebox vs ElevenLabs for local voice cloning, VoxCPM2 vs ElevenLabs for self-hosted TTS, and Breeze TTS 2's licensing.

FAQ

Q: Is Otter worth it for a small business?
A: For meeting-heavy teams, yes: $16.99 per user monthly or $8.33 annually for 1,200 recording minutes is the lowest full commercial plan here. Less suitable if you need many languages or verified accuracy.

Q: What is the cheapest transcription software?
A: It depends on volume. At 10 hours a month, Otter Pro works out near $1.70 per hour; TurboScribe's unlimited plan gets cheaper per hour as you upload more.

Q: Is there a free transcription tool that is good enough?
A: Otter's free plan offers 300 minutes a month with 30-minute conversations and three lifetime imports. Rev's free plan gives 45 English-only AI minutes. TurboScribe's allows three files a day of up to 30 minutes.

Q: Rev vs Otter: which is better for accuracy?
A: Rev, if you buy its human service: 99%+ accuracy is guaranteed from $1.99 per minute. Otter is AI-only and cheaper, and suits searchable meeting notes rather than records that need guaranteed accuracy.

Q: Which transcription software supports the most languages?
A: Happy Scribe, with 150+ AI languages and dialects. TurboScribe covers 98+ and Sonix 54+. Otter supports six.

Q: Can I keep audio on EU servers?
A: Happy Scribe lists EU data residency in Dublin, Ireland (independent comparison). For other vendors, check the current data-processing terms before you upload sensitive recordings.

Last verified: 10 November 2026. Vendors change plans often, so confirm the price on the vendor's page before you buy. Corrections are welcome and will be logged here.

This article was researched and drafted with AI assistance and reviewed against vendor pricing pages. Our editorial standards and corrections policy explain how.

OpenAI Business Hub for Small Business: Free in India

Shamuddin — Sun, 15 Nov 2026 04:43:15 +0000

Verdict: Claim your profile. OpenAI launched Businesses in ChatGPT in beta on November 10, 2026, and the Business Hub half of it is open to any eligible business located where ChatGPT is available, India included. Claiming costs Rs 0, takes roughly 30 to 40 minutes, and needs no paid plan. The Business Agent half is limited to eligible ecommerce and online service businesses in the United States, so do not pay for anything expecting agent access in India.

This is the OpenAI Business Hub for small business in practical terms: a free, verified profile that controls what ChatGPT can say about you, in a product OpenAI says 1.2 billion people use each week. OpenAI's own figure, not an independent count.

TL;DR

Business Hub opens from the ChatGPT sidebar and lets you claim, verify and edit a business profile at no cost.
Verification: phone, business email, domain or Shopify, then a Persona identity check.
A 28-day dashboard shows mentions, website clicks, profile views and discovery topics.
Business Agents (a chat assistant for your customers) and sponsored agents are US-only for now, with no date given for expansion.
ChatGPT Go (Rs 399/mo) and Plus (Rs 1,999/mo) are optional, unrelated to claiming.
Best fit: an India online service business with a website - design studio, agency or clinic selling consultations.

What does the OpenAI Business Hub for small business actually include?

Business Hub is the self-service side of the launch, documented in OpenAI's announcement and its product page. Once claimed, a profile holds your website, description, category, logo, cover photo, delivery policy and return policy. Businesses that sell products can add a product feed with pricing and availability. Verified profiles carry a blue check.

TechCrunch observed that the profiles greatly resemble Google Business Profile. That comparison is useful: treat this as a listing you maintain, not a product you configure once. OpenAI tried in-chat commerce earlier, with a commission fee and a checkout cart, and retired it in March 2026. This launch takes a different route through free profiles and agents.

What can an Indian business use today, and what is US-only?

Eligibility is split by feature. The table reflects OpenAI's launch documentation as of November 11, 2026.

Feature	India today	United States
Profile claim and verification	Yes	Yes
Insights dashboard	Yes	Yes
Product feed with pricing and availability	Yes	Yes
Business Agent chat	No	Yes, for eligible ecommerce and online service businesses
Sponsored agent via ChatGPT Ads	No	Yes

If you are weighing the advertising side, our note on ChatGPT Ads and the India ads manager covers where that stands. Digiday framed the free agent as the organic way into a format OpenAI also sells as advertising. OpenAI declined to share a timeline for wider rollout.

How do I claim my business profile in 30 minutes?

The steps follow OpenAI's setup guide. Worked example: an India-based design studio with its own website. Have your website login, a domain email and a phone handy.

Open Business Hub. Sign in to ChatGPT, then choose Business Hub from the sidebar (it also lives at chatgpt.com/business-hub). A free ChatGPT account is enough.
Start a claim. From the Business Hub sidebar, search for your business or begin a new claim, then enter your legal or trading name and website.
Choose a verification method. Pick phone, business email, website domain or Shopify. A studio with a custom domain usually finds domain verification quickest. Shopify merchants can connect the store from Business Hub or from the Shopify admin.
Complete the Persona identity check. This is required before you can publish profile changes, see insights or create an agent. Keep a government ID available and expect a few minutes here.
Correct the profile fields. Update description, category, logo, cover photo, and the delivery and return policies. A studio should state turnaround times and revision terms plainly, since those are the answers ChatGPT is most likely to repeat.
Read the insights panel. Check mentions, clicks, profile views and discovery topics - your baseline, rechecked monthly.

What do the insights tell me, and what do they hide?

The dashboard covers a 28-day window. Mentions count ChatGPT references to your business. Discovery topics show themes across conversations where your site was cited. Data only appears above privacy thresholds, so a very small business may see sparse or empty panels for a while. OpenAI states it does not share users' conversations or agent chats with businesses, which means you will see counts and themes, not transcripts.

That limits what you can attribute. A click is not a lead, and a mention is not a recommendation. Use the panel to spot gaps, for example a discovery topic you do not serve or a service ChatGPT never mentions, and then fix the page content behind it. ContentGrip made the case that keeping this information accurate becomes an ongoing marketing task, which matches how the tool behaves.

How much does it cost in rupees?

Claiming the profile is free. OpenAI's help documentation does not list a subscription or an Ads account as a requirement, and no ChatGPT Enterprise agreement is needed.

If you want ChatGPT for daily work such as drafting proposals or summarising client briefs, the optional plans in India are priced as follows:

ChatGPT Go: Rs 399 a month, GST included. The India promotion that made it free for eligible users ended on November 7, 2026.
ChatGPT Plus: Rs 1,999 a month, GST included.

Compare assistants in our 2026 guide; the TurboAgents price breakdown covers Indian-priced marketing platforms.

What are the risks and limits?

Three matter. First, the product is in beta, so fields, eligibility and verification options may change. Second, a profile is only as good as your website: a vague pricing page defeats a verified badge. Third, the Business Agents policy, updated November 10, 2026, bars passing an agent off as a person or independent advisor, inventing offers, prices, reviews or guarantees, and pushing a sale after a refusal. Those rules will apply to you once agents reach India, so draft your policies now.

New to agents? Our agentic AI versus generative AI explainer covers what a Business Agent will and will not do.

FAQ

Q: Is the OpenAI Business Hub free for Indian businesses?
A: Yes. Claiming in Business Hub costs Rs 0 and needs no paid plan, Ads account or enterprise agreement, per OpenAI's documentation.

Q: Can I create a Business Agent for my Indian business?
A: Not yet. At launch they are limited to eligible ecommerce and online service businesses in the United States; OpenAI has given no date for expansion.

Q: What verification do I need to publish changes?
A: Phone, business email, website domain or Shopify, then a Persona identity check - without it you cannot publish changes, view insights or create an agent.

Q: Do I need ChatGPT Go or Plus to use Business Hub?
A: No. Go (Rs 399) and Plus (Rs 1,999) are optional daily-work tools; neither is required to claim.

Claude Code 2.1.321: Fable 5.2 vs 5.1, 20% Cheaper

Shamuddin — Sat, 14 Nov 2026 04:44:57 +0000

Claude Code 2.1.321, published on GitHub on 9 November 2026, makes Claude Fable 5.2 (claude-fable-5-2) the default Fable model, priced at $8 per million input tokens and $40 per million output tokens with a 1M-token context window (release notes). Fable 5.1 cost $10 and $50 for the same units (Classmethod), so the new model is 20% cheaper on input, output and cache reads. The verdict: if you already use Fable in Claude Code, 5.2 wins over 5.1 on price, speed and Anthropic's published benchmarks, and the upgrade is close to automatic. Anthropic's own guidance still points to Opus 5.5 as the daily driver for most coding work.

TL;DR

Claude Code 2.1.321 adds Fable 5.2 and makes it the default Fable model; saved Fable 5.1 picks move to the fable alias.
Unit prices fall from $10/$50/$0.25 to $8/$40/$0.20 per million tokens (input, output, cache reads), a 20% cut across the board (release notes; Classmethod).
Anthropic says Fable 5.2 generates output up to 30% faster than Fable 5.1 (Anthropic).
The release also fixes four sandbox, permission and secret-handling bugs that matter if you run agents unattended.
You can revert with /model claude-fable-5-1. Mythos 5.2 is a separate, restricted variant and is not needed for normal development.

What changed in Claude Code 2.1.321?

The headline item is the model swap. The release notes describe Fable 5.2 as the new default Fable model, with 1M context and cache reads at $0.20 per million tokens. Anthropic's platform documentation lists a 128K maximum output, always-on adaptive thinking and a June 2026 knowledge cutoff (model overview).

Migration is automatic in most cases. Saved Fable 5.1 selections move to the fable alias immediately on Pro and Max plans. On Team and Enterprise plans the move happens once your organisation allows the new model, so an admin setting may delay it.

The release also adds a --settings <file> option to claude remote-control, which applies one settings file to every session that command starts. That is useful if you manage permission rules centrally.

How does Fable 5.2 compare with Fable 5.1 and Opus 5.5?

The figures below come from Anthropic's launch announcement and release notes. Benchmark scores are Anthropic's own, run at the effort levels it chose, so treat them as vendor claims rather than independent results.

	Fable 5.2	Fable 5.1	Opus 5.5
Input / output per Mtok	$8 / $40	$10 / $50	Not stated in the release material
Cache reads per Mtok	$0.20	$0.25	Not stated
FrontierCode 2.0 (Xhigh)	49.1%	32.4%	45.5%
FrontierSWE v2	69.9%	56.3%	62.3%
OSWorld 2.1 (offline set, partial)	89.3%	83.5%	87.2%
Remote Labor Index	41.3% (High effort)	17.9%	21.3%

Sources: Anthropic announcement for benchmarks and Fable 5.2 pricing; Classmethod for Fable 5.1 pricing.

Two readings are worth making. First, the gap over Fable 5.1 is large on agentic coding: 49.1% against 32.4% on FrontierCode 2.0. Second, Opus 5.5 is close behind on coding tasks. Anthropic's accuracy-versus-cost data shows Fable 5.2 at High effort scoring about level with Opus 5.5's best FrontierCode 2.0 result (45.3% against 45.5% (Anthropic announcement)) for roughly half the cost per task. If you pay per token, that cost-per-task point matters more than the headline score. For a wider look at how model pricing is shifting, see our Claude Sonnet price comparison against GLM 5.3.

How much does the 20% price cut save in practice?

The saving is simple arithmetic on the published rates, so no benchmark hype is needed. Take an illustrative session that reads 2 million input tokens and writes 200,000 output tokens (figures chosen for easy arithmetic, not measured):

Fable 5.1: 2 x $10 + 0.2 x $50 = $30.00 (Classmethod rate card)
Fable 5.2: 2 x $8 + 0.2 x $40 = $24.00 (release notes)

That is $6 saved per session, or 20%. Real sessions vary, because cache reads, thinking tokens and tool output all change the mix, but the ratio holds as long as the token mix is the same. A faster model may also finish sooner, though you should measure that on your own repositories before counting on it.

If you want a larger reduction, the savings from structural changes can exceed a price cut. Our piece on cheaper subagent status updates in Claude Code covers one such approach.

How do I check and switch models after updating?

Follow these steps:

Update Claude Code to version 2.1.321 or later using your usual install method.
Run claude --version and confirm it reports 2.1.321 or higher.
Start a session and run /model to see which model is active. On Pro and Max plans a previous Fable 5.1 choice should now show the fable alias.
To go back to the older model, run /model claude-fable-5-1.
Run /fast to check whether fast mode is on. Anthropic lists fast mode at up to 2.5x faster token generation, at $16 per million input tokens and $80 per million output tokens, which is double the standard rate.

Fast mode is a sensible choice for interactive work where waiting is the cost. For long unattended runs, standard mode is usually the cheaper pick.

Which security fixes in Claude Code 2.1.321 matter for unattended agents?

Four fixes in the release notes affect anyone who lets Claude Code run commands without close supervision:

MCP OAuth requests no longer send secrets, such as refresh tokens, to a different host when a server redirects them.
A read-only Bash command that could run outside the sandbox without a permission prompt now prompts correctly.
Bash permission checks no longer judge a file write from the wrong folder and skip the prompt.
Deny rules and protected-file prompts now apply to the file that a cp or mv into a directory actually writes.

The fourth is the easiest to underestimate. A deny rule that protects .env is only useful if copying or moving a file into a directory cannot get around it. These fixes follow a security batch in versions 2.1.317 to 2.1.319 a few days earlier (AI Coding Guide), so updating promptly is the practical response. If your agents handle credentials, our guide to Bitwarden versus 1Password for Claude Code secrets covers how to keep them out of reach.

What are the limitations of this upgrade?

Benchmarks are vendor-reported, and scores at different effort levels are not directly comparable. The Fable 5.2 figure for the Remote Labor Index uses High effort while the coding result uses Xhigh. The cost-per-task claim comes from Anthropic's own charts. No independent test of Fable 5.2 on your codebase exists yet, so run a few representative tasks before changing team-wide defaults.

If you work across editors, the picture differs by tool. Our notes on GitHub Copilot in Xcode 27.3 and Cursor's hosted web apps and cloud agents cover alternatives with their own model and pricing arrangements.

Why we covered this release

Of the 694 AI and developer-tooling keywords we priced for the keyword corpus, 72 cleared a winnable bar and 20 of those carry zero keyword difficulty — meaning no established competitor holds the result set (n=694 corpus, filtered to 72 winnable; measured 2026-11-09). Release-day model comparisons like this one sit in exactly that gap: the pricing and benchmark facts are hours old, and nobody has organised them for buyers yet.

FAQ

Q: Is Fable 5.2 the default model in Claude Code 2.1.321?

A: Yes, it is the default Fable model. Saved Fable 5.1 picks move to the fable alias immediately on Pro and Max plans, and on Team and Enterprise plans once your organisation allows it.

Q: How much cheaper is Fable 5.2 than Fable 5.1?

A: It is 20% cheaper. Fable 5.2 costs $8 input, $40 output and $0.20 cache reads per million tokens, against $10, $50 and $0.25 for Fable 5.1 (release notes; Classmethod).

Q: How do I switch back to Fable 5.1?

A: Run /model claude-fable-5-1 inside Claude Code. You can run /model with no argument to confirm which model is currently active.

Q: Do I need Mythos 5.2 verification to use Fable 5.2?

A: No. Fable 5.2 is available to all users. Mythos 5.2 is the same underlying model with some safeguards removed, and it is restricted to organisations in Anthropic's Life Sciences and Cyber Verification Programs.

Q: Is fast mode worth it for Fable 5.2?

A: It depends on what you are optimising. Fast mode generates tokens up to 2.5x faster at twice the standard price ($16 input, $80 output per million tokens (Anthropic)), so it suits interactive sessions where waiting costs more than tokens, but not long unattended runs.

Anthropic IPO vs SpaceX: Smaller Raise, Tighter Control

Shamuddin — Sat, 14 Nov 2026 04:44:36 +0000

The Anthropic IPO roadshow opened on 10 November 2026, offering 825 million Class A shares at $70 to $78 each under the Nasdaq ticker ANT, according to the company's press release. Trading could begin around 19 November, CNBC reports. Our verdict: SpaceX still wins on size, but Anthropic wins on retail access, while ANT buyers get far less say over the company than the valuation suggests.

At the top of the range, the offering raises up to $64.35 billion and implies a fully diluted valuation of up to $1.404 trillion, per TechCrunch.

TL;DR

The Anthropic IPO offers 825 million shares at $70 to $78, listing on Nasdaq as ANT (press release).
The top of the range raises $64.35 billion, below the $85.7 billion SpaceX raised in June (TechCrunch).
A separate $26 billion private placement goes to seven suppliers, including Google and Nvidia.
Up to 20% of the offering is reserved for retail buyers, which is high by IPO standards (CNBC).
Public shareholders sit behind a Class F share with 50.1% of the votes and a trust that elects a board majority.

What exactly is the Anthropic IPO offering?

The press release splits the 825 million Class A shares into 801,351,354 sold by Anthropic and 23,648,646 sold by existing stockholders. Anthropic receives no proceeds from the stockholder portion. Underwriters hold a 30-day option on a further 123,750,000 shares (press release).

At the ends of the range, 825 million shares at $70 is $57.75 billion and at $78 is $64.35 billion (TechCrunch). The listing is on the Nasdaq Global Select Market. The company filed an amended S-1 with the SEC alongside the launch.

Seven founders and senior staff are selling about 3.4 million shares each, including Dario Amodei and Daniela Amodei. At the $74 midpoint that is roughly $250 million apiece, TechCrunch reports.

Why are seven suppliers buying $26 billion of stock?

Alongside the public offering, Anthropic is selling non-voting Class N stock at the IPO price to seven suppliers, TechCrunch reports. The split below comes from CNBC and Yahoo Finance coverage of the filing, and is approximate for the smaller names.

Supplier	Reported placement
Google	$10 billion
Nvidia	$10 billion
AMD	about $2.5 billion
Broadcom	about $1 billion
Micron	about $1 billion
Microsoft	about $1 billion
TSMC	about $500 million
Total	$26 billion

The logic is circular but coherent. Anthropic has roughly $900.9 billion of decade-long spending commitments, mostly compute and infrastructure, with about $70.7 billion due within 12 months, according to the SEC filing. The suppliers selling it chips, memory and cloud capacity are now also owners. Including the placement and the over-allotment, estimated net proceeds run from $84.6 billion to $93.7 billion (TechCrunch).

For the memory and foundry side of that chain, see our comparison of Micron and Tata in India's semiconductor market, which covers TSMC's position.

What do ANT shareholders actually control?

Very little. Anthropic is a public benefit corporation. A single Class F share carries 50.1% of the voting power, and the Long-Term Benefit Trust elects a majority of the board, TechCrunch reports. Class A buyers get economic exposure and a minority vote. Class N holders, including Google and Nvidia, have no voting rights at all.

That is a legitimate design choice for a company that says safety should outrank short-term returns, but it is a real tradeoff. If you disagree with the board, the practical remedy is to sell.

How does the Anthropic IPO compare with SpaceX?

	Anthropic (ANT)	SpaceX
Listing	Nasdaq, expected around 19 November 2026	June 2026
Raise	$57.75 billion to $64.35 billion	$85.7 billion
Valuation	Up to $1.404 trillion fully diluted	Market cap just under $2 trillion after listing
Structure	PBC, Class F votes, trust-elected board majority	Not covered here

All figures are from TechCrunch and CNBC. SpaceX was the larger raise, so it keeps the record for proceeds. Anthropic's offer is distinctive for who can buy and how little control they get.

Can retail investors buy ANT, and what about India?

Up to 20% of the offering is reserved for retail, CNBC reports. An Australian retail window opened on 11 November and closes at 5pm AEDT on 17 November, which matters for readers in Asia-Pacific time zones.

India has no direct allocation route that we can confirm. Indian residents who invest in US stocks usually do so through platforms such as Vested or INDmoney, under the RBI's overseas remittance rules. Check whether your platform offers IPO allocations or only trading after the listing, because many offer the second only.

Anthropic's India links are growing in other ways, as our report on the Anthropic licensing deal with ANI shows. A listing does not by itself change Claude API pricing, and nothing in the roadshow materials we cite sets any.

What is the business behind the valuation?

Run-rate revenue reached about $81 billion by the end of October, up around $10 billion in a month (TechCrunch). For scale against a rival's advertising line, see our piece on ChatGPT's ads reaching a $1 billion run rate.

The timing also coincides with model news. The company's Fable 5.2 and Mythos 5.2 releases on 9 November retook first place on Artificial Analysis, per the AIsa daily (AI Daily, November 11, 2026). Our coverage of the White House, Moonshot and distillation dispute gives the policy background. The listing itself is also on our November and December AI events calendar.

FAQ

Q: Can retail investors buy ANT shares in the IPO?

A: Up to 20% of the offering is reserved for retail buyers, according to CNBC. Access depends on your broker, and allocations are not guaranteed.

Q: Can Indian investors buy ANT?

A: Most Indian investors would need a platform with US market access, such as Vested or INDmoney, and should confirm whether it offers IPO allocations or only post-listing trading.

Q: What happens around 18 and 19 November?

A: Trading on Nasdaq could begin around 19 November, CNBC reports. The final price is usually set after the roadshow, so the range may move.

Q: Why is the valuation below earlier reports of a 2 trillion dollar target, per Fortune?

A: Earlier reports of a roughly $2 trillion target, per Fortune, did not materialise. The range was trimmed from about 867 million shares at $70 to $80 the day before launch, per TechCrunch. The May 2026 fundraise valued the company at $965 billion.

Q: What are the main risks for buyers?

A: Governance and cost. The Class F share holds 50.1% of votes, a trust elects the board majority, and the filing shows about $900.9 billion of commitments against roughly $81 billion of run-rate revenue.

Q: Is the $26 billion supplier placement unusual?

A: Yes, in size and structure. Sellers of chips and cloud capacity are buying non-voting stock at the IPO price, which ties their revenue to Anthropic's growth (TechCrunch).

This article was drafted with AI assistance and verified against the primary sources linked above. See how we work.

LongCat-2.5 Open Weights vs Kimi K3.1 Pro: MIT Wins

Shamuddin — Sat, 14 Nov 2026 04:44:16 +0000

On 10 November 2026 Meituan published the LongCat-2.5 open weights on Hugging Face under the MIT licence: a 1.6 trillion parameter mixture-of-experts model with about 48 billion parameters active per token, in a 3.56 TB repository. On licence terms it beats Kimi K3.1 Pro and Baidu's ERNIE 5.5 Flash Base, because neither revenue gates nor territory carve-outs apply. For most Indian developers and teams the sensible order is to rent the hosted API first, at roughly ₹28 per million input tokens (pricing), and self-host only if you already operate trillion-parameter MoE serving.

TL;DR

Meituan released the weights one week after launching LongCat-2.5 as a hosted service on 1 November 2026.
The licence is plain MIT. It grants no rights to Meituan trademarks or patents, which matters for legal review.
The checkpoint is 3.56 TB across 208 safetensors shards. Meituan's tested serving path targets NPUs, not GPUs.
Hosted pricing is a limited-time discount: $0.30 input and $1.20 output per million tokens, about ₹28 and ₹114.
Benchmarks are mostly in-house and flagged by Meituan as indicative only.
Verdict: rent first, run later. MIT makes it the lowest-risk large open-weights base to build on this month.

What changed from LongCat-2.0?

LongCat-2.5 builds on LongCat-2.0 and adds native visual input alongside agentic coding. According to the model card, vision comes from LongCat-Next, which turns native-resolution visual features into discrete tokens aligned with text. It is a hybrid model: thinking and non-thinking modes are switched per request with chat_template_kwargs: {enable_thinking: ...} in the OpenAI-compatible API.

Benchmark (model card, in-house)	LongCat-2.0	LongCat-2.5
Terminal Bench v2.1	70.8	88.4
Terminal Bench v4.0	3.0	38.5
DeepSWE v1.1	9.7	64.6
BrowseComp	79.9	91.9
AutomationBench v1.0.6	15.8	51.8

Meituan states that unmarked scores are measured in-house and are indicative only, so treat the jumps as a direction rather than a guarantee. On the same card, Terminal Bench v2.1 puts Claude Opus 5 at 89.1 and DeepSeek-V4.1-Flash at 90.6, both cited from public reports, so LongCat-2.5 is close to the leaders rather than ahead. VentureBeat notes that the comparison set predates newer models, some of which post higher Terminal Bench v4.0 scores than LongCat's 38.5; even Claude Opus 5's cited 53.9 on that suite is well clear of the 38.5, per the model card comparison.

How does the MIT licence compare with Kimi K3.1 Pro and ERNIE?

The large open-weights releases this November now differ more by licence than by capability. The LongCat LICENSE file is 1.06 kB and reads "Copyright (c) 2026 Meituan", with the model card confirming the weights are MIT.

Model	Licence	Main constraint
LongCat-2.5 (Meituan)	MIT	None beyond attribution; no trademark or patent grant
Kimi K3.1 Pro (Moonshot)	Modified MIT	Separate agreement above a revenue threshold for model-as-a-service use; see our Kimi K3.1 Pro write-up
ERNIE 5.5 Flash Base (Baidu)	Custom	Excludes the EU, UK and South Korea; commercial use needs notice or authorisation by revenue band

The Kimi and ERNIE rows come from third-party write-ups of their licence files, so read the originals before committing a product to either. If you are choosing a base model for a company, MIT removes a whole category of legal questions. For a comparison of two other recent releases, see Tencent HY4 vs Mistral Large 4.

How do I download the LongCat-2.5 open weights?

The repository shows this command pattern:

hf download hf://meituan-longcat/LongCat-2.5

Expect 208 shards of roughly 17.2 GB each, plus an 8.12 MB index file. Check free disk before starting; the checkpoint alone is 3.56 TB, before any converted or quantised copies. The upgrade path from LongCat-2.0 is a straightforward swap of model path and chat-template flags, provided your serving stack supports the architecture.

That is the catch. Hugging Face lists no inference provider hosting the model, and Meituan's documented serving route is NPU platforms through SGLang-FluentLLM. GPU users should wait for community vLLM or SGLang support. Meituan reports W4A4 quantisation on expert weights, with 43.3% higher prefill throughput and 42.77% higher decode throughput than a W8A8 baseline, per the model card. Those figures describe Meituan's own stack, not what you will see on other hardware.

Should I rent the API or run LongCat-2.5 myself?

Rent first. The LongCat pricing page lists a limited-time discount of $0.30 per million uncached input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens. Using the RBI reference rate of 94.86 rupees per dollar from 3 November 2026, that is approximately ₹28, ₹0.57 and ₹114. All rupee figures here are approximate and will move with the exchange rate and the end of the discount.

A worked example: at the listed promo rates of $0.30 per million uncached input tokens and $1.20 per million output tokens (LongCat pricing), an agent pipeline using 10 million uncached input tokens and 2 million output tokens a day comes to $3.00 plus $2.40, about $5.40 a day, roughly ₹512 at those rates (pricing page), around ₹15,400 over 30 days. That is far below the cost of keeping a multi-node cluster busy for a 3.56 TB model.

Situation	Better choice
Prototype, evaluation, spiky traffic	Hosted API
Existing trillion-parameter MoE serving, NPU access	Self-host
Data must not leave your network	Self-host, once GPU support matures
Need a legally clean base to fine-tune or redistribute	Download the MIT weights

The hosted price is promotional, so rerun the numbers when it ends. Also note the model is listed on Vercel's AI Gateway, per TLDR, which gives a second route to try it.

Why does this release matter for developers searching for open models?

Specific model names are where developers actually search, and they are a narrow target. We priced 694 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data. Only 72 (10.4%) cleared a winnable bar of 150-6,000 monthly searches, difficulty 20 or below, a genuine technical term, and at least three words. (n=694, measured 2026-11-09, keyword volume and difficulty data via DataForSEO). A concrete query such as "LongCat-2.5 open weights" is the kind of term that passes that bar, because it names a real artefact. If you are weighing coding-agent tooling to pair with it, our guide to open-source agent platforms for coding is a reasonable next read.

FAQ

Q: Is LongCat-2.5 really open source?
A: The weights are published under the MIT licence, which is permissive. Training data and full training code are not part of the release, so "open weights" is the more accurate term.

Q: Can I use LongCat-2.5 commercially in India?
A: The MIT licence places no revenue or territory limits on commercial use. It does not grant rights to Meituan trademarks or patents, so do not use Meituan branding without permission.

Q: Can I run LongCat-2.5 on GPUs today?
A: Meituan's tested path is NPU-based through SGLang-FluentLLM. No public inference provider hosts it yet, so GPU users should wait for community adapter support.

Q: How big is the LongCat-2.5 download?
A: The repository is 3.56 TB across 208 safetensors shards of about 17.2 GB each, according to the Hugging Face page.

Q: Are the LongCat-2.5 benchmark scores reliable?
A: Treat them as indicative. Meituan labels unmarked scores as in-house measurements, and the comparison models are not the latest releases.

Q: Are smaller open models worth considering instead?
A: For narrow tasks, often yes. A specialised model such as PLaMo 3 Translate runs offline and fits on far more modest hardware, and Arete 27B shows what a mid-sized post-train can do.

Andhra Pradesh AI Data Centres: 25 Acres for Reliance

Shamuddin — Sat, 14 Nov 2026 04:43:56 +0000

On November 10, 2026, the Andhra Pradesh Cabinet approved about 25 acres for Reliance Industries to build a gigascale AI data centre at Bhogapuram, cleared a Rs 31,387 crore, 512 MW AI data centre near Visakhapatnam, and approved a Rs 570.20 crore Epack equipment factory at Rambilli expected to employ 2,565 people, according to The Hindu's report on the meeting. Of the three, the Reliance land allocation is the newest development, while the Vizag project and the Epack plant show the state moving on both the facility and the hardware that goes into it. The practical reading: Andhra Pradesh is now the clearest non-Gujarat location for Reliance's AI infrastructure plans, though none of these decisions means a facility is built or powered yet.

Last verified: 2026-11-11

TL;DR

The Cabinet, chaired by Chief Minister N. Chandrababu Naidu at the Secretariat in Velagapudi, allocated about 25 acres to Reliance Industries at Bhogapuram, Vizianagaram district.
It also cleared AM Intelligence's Rs 31,387 crore, 512 MW green AI data centre at Amanam, Visakhapatnam district, with about 1,000 jobs expected (The Hindu).
Epack Data Centre Solutions got approval for a Rs 570.20 crore unit at Rambilli, Anakapalli district, for AI and hyperscale data centre equipment.
The Rs 31,387 crore figure works out to roughly Rs 61.3 crore per MW, a benchmark for judging other proposals (The Hindu).
Land and cabinet approvals are early steps. Power, grid connections and commissioning dates were not part of the reports reviewed.

What did the Andhra Pradesh Cabinet approve for AI data centres?

The Cabinet cleared more than 30 agenda items on November 10, 2026, per Deccan Chronicle. Three relate directly to AI infrastructure. The Times of India puts the total of investment proposals cleared at more than Rs 37,000 crore, which also includes Carbon Cosmo Advanced Materials' Rs 4,000 crore battery materials project at Naidupeta and Dixon Technologies' Rs 700 crore electronics project in Tirupati.

Project	Company	Location	Size	Investment	Jobs
Gigascale AI data centre	Reliance Industries	Bhogapuram, Vizianagaram	About 25 acres (land)	Not stated in reports reviewed	Not stated
Green integrated AI data centre	AM Intelligence	Amanam, Bheemunipatnam mandal, Visakhapatnam	512 MW	Rs 31,387 crore	About 1,000
Data centre equipment factory	Epack Data Centre Solutions	Rambilli, Anakapalli	AI and hyperscale equipment unit	Rs 570.20 crore	2,565

Sources: The Hindu and Hans India. The New Indian Express names the Visakhapatnam developer as AM Intelligence Labs Pvt Ltd.

What does 25 acres at Bhogapuram mean for Reliance?

The allocation is a land step, not a capacity announcement. Deccan Chronicle names APIIC, the state industrial infrastructure agency, as the allotting body. The reports reviewed give no megawatt target, investment figure or timeline for the Bhogapuram site; any estimate of its eventual size would be a guess.

What gives the allocation weight is Reliance's stated direction. At the India AI Impact Summit in February 2026, Mukesh Ambani described plans for multi-gigawatt AI-ready data centres in Jamnagar, with more than 120 MW online in the second half of 2026 and up to 10 GW of green power earmarked, as DD News reported. Bhogapuram extends that plan to the east coast. Within the policy record covered here, it is the first named Reliance data centre allocation outside Gujarat.

How much does a 512 MW AI data centre cost per megawatt?

Dividing Rs 31,387 crore by 512 MW gives about Rs 61.3 crore per MW. The inputs come from The Hindu; the division is our analysis of the approved figures.

Use it with care: the headline investment figure divided by headline capacity may fold in land, power infrastructure and phased spending that other proposals count differently. If another Indian proposal claims similar capacity at a very different per-MW cost, ask what each figure includes.

The AM Intelligence project is not new to the state record. An order in August 2026 had already approved the same 512 MW project with a completion deadline of two years and four months, and we covered it in our Vizag AI data centre explainer. The November 10 clearance advances it rather than introducing it.

Why does the Epack plant at Rambilli matter?

It addresses the supply side. Most coverage of Indian AI infrastructure counts facilities and megawatts; the Rs 570.20 crore Epack unit is about the equipment that goes into those facilities. Per the approval, it will make AI and hyperscale data centre equipment in Anakapalli district and create jobs for 2,565 people (The Hindu).

The jobs-to-investment ratio is also worth noting: 2,565 jobs against Rs 570.20 crore is a manufacturing profile, whereas the 512 MW AM Intelligence campus lists about 1,000 jobs against Rs 31,387 crore (Hans India). Data centres are capital-heavy and staff-light once built; factories are the opposite. The tradeoff: equipment makers depend on data centre demand actually materialising.

What else did the Cabinet decide on AI and technology policy?

The Cabinet approved the Andhra Pradesh Robotics and Physical AI Policy 2026-36 and the AVGC-XR Policy 2026-31, and cleared amendments to the Semiconductor and Display Fab Policy 2024-29, per Hans India. It also authorised the IT Department to examine the Adani Group's proposed cloud hub, according to Telugu360. That is an examination mandate, not an approval.

For readers tracking the demand side of this build-out, our pieces on IndiaAI foundation models and DEPA-based AI training in sealed rooms cover the workloads such capacity is meant to serve. For a different kind of applied AI deployment in India, see Indian Railways' AI camera rollout.

What should readers watch next?

Whether Reliance discloses capacity (MW) and investment for Bhogapuram.
Power sourcing and grid arrangements for the 512 MW campus, since the project is described as green.
Construction milestones against the two-year-four-month deadline set in August 2026.
Whether Epack's factory start dates and hiring match the 2,565-job figure in The Hindu's report.
Any outcome from the IT Department's examination of the Adani cloud hub.

FAQ

Q: How much land did Andhra Pradesh allocate to Reliance for an AI data centre?
A: The Cabinet approved about 25 acres at Bhogapuram in Vizianagaram district for a gigascale AI data centre (The Hindu). The allotting agency is APIIC, and no investment or capacity figure for the site appeared in the reports reviewed.

Q: How large is the AM Intelligence data centre near Visakhapatnam?
A: It is a 512 MW green integrated AI data centre at Amanam in Bheemunipatnam mandal, with a Rs 31,387 crore investment and about 1,000 expected jobs (Hans India). That is roughly Rs 61.3 crore per MW on a simple division.

Q: What will the Epack unit at Rambilli make?
A: Epack Data Centre Solutions will manufacture AI and hyperscale data centre equipment in Anakapalli district. The approved investment is Rs 570.20 crore, and employment for 2,565 people is expected (Hans India).

Q: When did the Andhra Pradesh Cabinet make these decisions?
A: The Cabinet met on Tuesday, November 10, 2026, chaired by Chief Minister N. Chandrababu Naidu at the Secretariat in Velagapudi, and cleared more than 30 agenda items (per Deccan Chronicle).

Q: Does cabinet approval mean the data centres are operational?
A: No. Cabinet clearance and land allocation come before construction, power connection and commissioning. The AM Intelligence project has a stated completion deadline of two years and four months from its August 2026 order.

Corrections log

No corrections to date. Figures reflect press reports of the November 10, 2026 Cabinet meeting; where a number was not disclosed, we have said so.

This article was produced with AI assistance and edited against the cited primary and press sources.

Gemini 4 Carbon in GitHub Copilot vs Argon: Test by Dec 1

Shamuddin — Sat, 14 Nov 2026 04:43:36 +0000

Gemini 4 Carbon GitHub Copilot support became generally available on November 10, 2026, according to the GitHub changelog. It is billed per token at $2.00 per million input tokens and $10.00 per million output tokens until November 30, then $4.00 and $20.00 from December 1 (GitHub Docs). My verdict: use Carbon in Copilot for a short trial on your hardest agentic tasks while the introductory rate lasts, keep a faster model such as Gemini 4 Argon for everyday work, and budget at the doubled rate.

TL;DR

Carbon is available in Copilot Pro+, Max, Business and Enterprise; Free and Pro are not on the list.
Billing is per token at provider list price, not a premium-request multiplier. One GitHub AI credit equals $0.01 (GitHub Docs).
Introductory rates end after November 30, 2026. From December 1 every token price doubles.
A typical agentic task in my worked example costs $0.444 now and $0.888 afterwards, on the unit prices in the GitHub billing reference and Google's launch post.
Google itself says Argon is the better fit for most software engineering. Carbon is for the difficult cases.
Record the results and decide before the price changes.

Who gets Gemini 4 Carbon in GitHub Copilot?

GitHub lists Copilot Pro+, Max, Business and Enterprise as the plans with access. Copilot Free and Copilot Pro are excluded, so a Pro subscriber will not see the model however long they wait (GitHub changelog).

On Business and Enterprise, admins control access through the model policy: new models switch on automatically under default enablement unless an admin turns them off, so check the policy before announcing anything - developers may already have the model in their pickers.

The changelog names ten surfaces: VS Code, Visual Studio, Copilot CLI, the Copilot coding agent, the GitHub Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse. Rollout is gradual - one editor may get it before another. Xcode users, start with our Copilot in Xcode setup guide.

How is Carbon billed compared with premium requests?

Earlier Copilot models were described with premium-request multipliers. Carbon is different: GitHub bills it at provider list pricing under usage-based billing, so you pay for the tokens each task consumes. Cost now depends on how long the model thinks and how much context the agent sends, not on how many prompts you type.

This matters because Google describes Carbon as a model that uses more tokens to explore thoroughly, reason deeply and run extended verification loops. LiteLLM's pricing notes add that thinking tokens bill as output and Carbon supports implicit caching only. Output is the expensive side of the bill, so a model that reasons at length costs more per task than the unit price suggests.

Does Gemini 4 Carbon GitHub Copilot pricing change on December 1?

Yes. The rates per million tokens (GitHub; Google):

Token type	Through Nov 30, 2026	From Dec 1, 2026
Input	$2.00	$4.00
Cached input	$0.10	$0.20
Output	$10.00	$20.00

The figures below are my own arithmetic on those unit prices; they ignore extra thinking tokens, so treat them as a floor.

Profile	Tokens	Cost now	Cost from Dec 1
A: single chat prompt	4K in, 1K out	$0.018 per prompt ($18.00 per 1,000)	$0.036 per prompt ($36.00 per 1,000)
B: agentic coding task	60K in, 240K cached, 30K out	$0.444 per task ($44.40 per 100)	$0.888 per task ($88.80 per 100)
C: long-horizon refactor	800K in, 100K out	$2.60 per task	$5.20 per task

In Profile B the breakdown at the introductory token prices is 60,000 fresh input tokens at $2.00 per million = $0.12, 240,000 cached tokens at $0.10 per million = $0.024, and 30,000 output tokens at $10.00 per million = $0.30. Output is about two thirds of the bill even though it is the smallest token count. At one credit per cent, Profile B is about 44 credits now and 89 after the change (GitHub Docs).

Why we covered this

We priced 694 keywords (n=694, measured 2026-11-09) in the AI and developer-tooling space using DataForSEO volume and difficulty data. Only 72 (10.4%) cleared a winnable bar of 150-6,000 monthly searches, difficulty 20 or below, a genuine technical term, and at least three words. This article sits in that bar with an expiry date: December 1.

How do I select Carbon in Copilot and test it properly?

Update your editor or CLI to the current release, since rollout is gradual.
Open the model picker in Copilot Chat or the agent view on a supported surface and look for Gemini 4 Carbon.
If it is missing, confirm your plan (Pro+, Max, Business or Enterprise) and, on a managed plan, ask your admin whether the model policy allows it.
Pick three or four tasks your current model handles badly: a multi-file refactor, a flaky test hunt, a security review of a PR.
Run each once with Carbon and once with your usual model; record credits spent, correction rounds and whether the change merged without edits.

Cost per accepted change is the number to compare. A model that costs twice as much per run but needs one attempt instead of three can be the cheaper option, and the reverse is equally likely.

Is Carbon better than Argon for coding?

For hard problems, possibly. For everyday work, Google's own guidance points elsewhere: its launch post calls Argon the ideal choice for the majority of software engineering and knowledge work.

Google reports Carbon first on CursorBench 4.0, Zapier's AutomationBench and Perplexity's WANDR at more than 2x lower cost per task than the next best external model, and second on the Multimodal Knowledge Agent benchmark (Google); it also claims a 0% attack success rate on Gray Swan's indirect prompt-injection benchmark. These are vendor-reported results I have not reproduced. For a head-to-head against Anthropic's flagship, see our Opus 5.5 vs Gemini 4 Carbon comparison.

Outside Copilot, Carbon is also in Google Antigravity for AI Ultra subscribers, with a /swarm preview splitting one prompt across several agents (Android Authority). Our Antigravity, Cursor and Claude Code comparison covers that tooling decision, the Claude Code pricing breakdown weighs flat plans against metered tokens, and security teams can see the GPT-6 cyber benchmark and pricing review.

What is the verdict on the buy window?

Carbon wins for hard, multi-step agentic work where a wrong first attempt is expensive. Argon, or whichever fast model you already use, wins for completions, small edits and quick questions, where Carbon's extra reasoning adds cost without much benefit.

Before November 30, run the trial above on a small set of real tasks while tokens cost half of what they will afterwards. Decide which task types justify Carbon. Afterwards, budget every Carbon task at the December 1 rates and cap its use with usage controls if your plan offers them.

The limitation is that the saving is modest in absolute terms: the Profile B trial is $44.40 per 100 tasks at the introductory rate versus $88.80 on the doubled rate (GitHub Docs). The point is not the discount, it is learning your real token consumption before the doubled price makes a mistake twice as costly.

FAQ

Q: Who can use Gemini 4 Carbon in GitHub Copilot?

A: Copilot Pro+, Max, Business and Enterprise users. Copilot Free and Copilot Pro are not included. Business and Enterprise admins manage access through the model policy.

Q: How much does Gemini 4 Carbon cost in Copilot?

A: It is billed per token at provider list price (GitHub Docs). Until November 30, 2026 that is $2.00 per million input tokens, $0.10 per million cached input tokens and $10.00 per million output tokens. One AI credit equals $0.01 (GitHub Docs).

Q: What happens to the price on December 1?

A: Every rate doubles to $4.00 input, $0.20 cached input and $20.00 output per million tokens (GitHub Docs); a task costing $0.444 now would cost $0.888 on the same token counts, per the worked example above.

Q: Is Carbon worth it compared with Argon?

A: Worth testing on hard agentic tasks, where Google positions it as strongest. For most day-to-day coding, Google recommends Argon, and Carbon's heavier reasoning will use more tokens than you need.

Barracuda ESG Exploited: Patch vs Admin UI Lockdown

Shamuddin — Sat, 14 Nov 2026 04:43:15 +0000

Barracuda ESG exploited is no longer a theoretical risk. Barracuda Networks confirmed on 9 November 2026 that attackers are chaining CVE-2026-116606 and CVE-2026-111560 against its Email Security Gateway, and CISA added both to its Known Exploited Vulnerabilities catalogue on 10 November with a federal deadline of 13 November. Of the two defences available, patching to firmware 9.4.0.029 or later wins; restricting the Admin UI is a same-hour stopgap that shrinks exposure without removing the flaw.

TL;DR

Two flaws chain to give root on the Barracuda Email Security Gateway (ESG): a pre-authentication authorisation bypass and an OS command injection.
Barracuda says exploitation is "affecting a small number of organizations" and has not published a count, victim names or indicators of compromise.
CISA added both CVEs to the KEV catalogue on 10 November 2026 and set a 13 November deadline for US federal civilian agencies, including forensic triage.
Verdict: patch first, restrict the Admin UI as a stopgap, then hunt for signs of earlier compromise. Patching alone does not remove an intruder.
No Indian victim is publicly named, but Indian organisations running the appliance carry a six-hour CERT-In reporting duty if they find exploitation.

Is the Barracuda ESG exploited in the wild, and by what chain?

Yes. Barracuda's security advisory was published on 2 November and updated on 5 and 9 November. The 9 November update states that the company has identified exploitation activity affecting a small number of organisations. The Canadian Centre for Cyber Security also notes that open source reporting indicates CVE-2026-116606 is exploited in the wild.

The two flaws work as follows:

CVE-2026-116606 is a pre-authentication authorisation bypass with a CVSS score of 9.3. It affects only the ESG. Barracuda lists authentication as not required, but the Administrative UI must be reachable by the attacker.
CVE-2026-111560 is a command injection with a CVSS score of 8.6. It affects the ESG, Load Balancer ADC, Message Archiver, WAF Control Center, Web Application Firewall and Web Security Gateway, and it needs an admin login on its own.

Chained, the first removes the need for credentials and the second supplies code execution. CISA's KEV entries describe the mechanism at a high level: an unauthenticated attacker invokes the configuration-import function and restores a crafted backup archive, which alters the global configuration, and injected commands then run with root privileges when a crafted timezone value is processed. This article deliberately goes no further than that.

Patch or lock down the Admin UI: which defence wins?

Patching wins. The table compares the two on the points that matter during an active campaign.

Factor	Patch to 9.4.0.029+	Restrict Admin UI to internal networks
Removes the vulnerability	Yes	No
Stops the no-login attack from the internet	Yes	Yes, if the UI was the only exposure
Protects against an attacker already inside the network	Yes	No. Barracuda says less exposed devices remain vulnerable
Evicts an existing intruder	No	No
Speed	Needs firmware update, two steps below 9.1	Can be done within minutes with a firewall rule

Barracuda's guidance covers both: keep the Admin UI off untrusted networks, verify the installed firmware version, and keep automatic patching on. If it was disabled, patch manually. Close the exposure first because it takes minutes; patch immediately afterwards because only the patch removes the flaw.

Which patch IDs and firmware versions fix it?

For the ESG, Barracuda lists BNSF-40425 for firmware 9.4.017 and later, and BNSF-40430 for builds older than 9.4.017. Appliances should end on 9.4.0.029 or later according to Barracuda's documentation. Units below 9.1 must update in two steps.

The other products are affected by CVE-2026-111560 only:

Load Balancer ADC and WAF: BNADC-16586
Message Archiver: BNMA-32468
WAF Control Center: BNWF-60060
Web Security Gateway: BNYF-24789

After updating, check the Current Installed Version field rather than assuming the job completed.

Does patching evict an attacker who is already in?

No. CISA flagged both entries for forensic triage, a statement that exploitation may predate your patch. The exposure window runs from the advisory on 2 November, and watchTowr Labs published the full chain on 5 November, so working knowledge of it has been public since then.

A sensible hunt over that period looks for:

Configuration restores nobody on your team performed
New admin accounts or API keys
Firmware patch events you did not schedule
Unexpected outbound connections from the appliance

Barracuda has released no indicators of compromise for the attacks it saw, so these checks are generic, a starting point rather than a guarantee. watchTowr's GitHub detection artefact generator tests susceptibility.

A quick external check: from outside your network, see whether the gateway answers on its admin interface, which commonly listens on port 8000. A timeout or refused connection is the result you want.

What do we still not know about the Barracuda ESG attacks?

Several things:

How many organisations were breached. "A small number" is the vendor's only wording.
Who is behind the activity, and whether any data was stolen.
Whether attackers actually reached root in the observed intrusions.
Whether any Indian organisation is affected.

The history explains why this matters. In May 2023 the same product line was mass-exploited through CVE-2023-2868, attributed by Mandiant to UNC4841, and Barracuda ended up replacing appliances rather than patching them. Nothing published so far says the 2026 campaign is of that scale.

What does the Barracuda ESG exploited alert mean for Indian organisations?

No Indian victim has been publicly named, and CERT-In has not issued a separate ESG notice at the time of writing. The obligation still applies. Under the CERT-In Directions of 28 April 2022, issued under section 70B(6) of the IT Act, service providers, intermediaries, data centres and body corporates must report cyber incidents within six hours of noticing them. Reports go to incident@cert-in.org.in or 1800-11-4949. An Indian organisation that finds exploitation of its Barracuda gateway is inside that window from the moment it notices.

How does this compare with other recent exploited edge flaws?

The pattern is familiar: a perimeter appliance, a short gap between patch and exploitation, and a KEV deadline. See our coverage of OpenSSH exploited 40 hours after the patch, the three Ivanti Connect Secure zero-days and their CISA deadline, the WordPress CVE-2026-119977 patch-first decision and the Synology C3rb3r ransomware campaign against outdated DSM. The lesson repeats: unpatched edge devices are where attackers look first.

FAQ

Q: What are the CVEs in the Barracuda ESG exploited chain?
A: CVE-2026-116606 (CVSS 9.3) is a pre-authentication authorisation bypass affecting only the ESG, and CVE-2026-111560 (CVSS 8.6) is a command injection affecting several Barracuda products. Together they give root access on the ESG.

Q: What is the CISA deadline for the Barracuda flaws?
A: 13 November 2026, per CISA's 10 November KEV addition, for US federal civilian agencies to patch and complete forensic triage. Everyone else should treat it as a same-day priority.

Q: Which firmware version fixes the Barracuda ESG vulnerabilities?
A: Appliances should be on firmware 9.4.0.029 or later, applied through patch BNSF-40425 or BNSF-40430 depending on the current build. Units below 9.1 need a two-step update.

Q: Is restricting the Admin UI enough to stay safe?
A: No. It cuts internet exposure to the no-login attack, but the flaw remains and an inside attacker can still reach the UI. Patch as well and hunt for earlier compromise.

Q: Do I need to report Barracuda ESG exploitation in India?
A: Yes, if you notice it. CERT-In's 2022 Directions require reporting of cyber incidents within six hours to incident@cert-in.org.in or 1800-11-4949.