DEV Community

Sh Raj
Sh Raj

Posted on

Five Models in Thirteen Days: What November's AI Wave Actually Means for Developers

Between November 1 and November 13, 2026, three vendors shipped five text or multimodal models you can call by API. Anthropic released Claude Fable 5.2 and Mythos 5.2. OpenAI released GPT-6-Cyber. Google made Gemini 4 Carbon and Gemini 4 Hydrogen generally available. OpenAI also shipped two image models, and Google shipped a video model and a decision-making API.

That is a lot of launch posts to read in under two weeks. But the number that matters most is not five. It is one: the number of these models that most developers can actually use today, on a free tier, without an application process. That model is Gemini 4 Hydrogen, and it went generally available on November 12.

This is a map of the wave — what shipped, what it costs, the catches buried in the fine print, and the December deadlines that matter more to your production system than any of the launches.

The headline: Hydrogen is for everyone

Google's announcement of Gemini 4 Hydrogen, published November 12 on the Google blog by Koray Kavukcuoglu (SVP at Google DeepMind and Chief AI Architect at Google), positions it plainly: "bringing cutting-edge intelligence to everyone — with speed and efficiency for everyday coding and agentic workflows."

The facts from Google's own post:

  • Price: $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire 1 million-token context window.
  • Availability: all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity. It is rolling out globally, and it is the third Gemini 4 model — after Argon and Carbon — but the first in Search for all users rather than only paying subscribers.
  • Performance claims (Google-reported): Hydrogen surpasses Argon on the Artificial Analysis Intelligence Index at its default high thinking level (51 vs. 50), reaching 54 at its highest xhigh thinking setting, "while often completing tasks faster than models in its class." Treat vendor index scores as vendor claims — useful for orientation, not a substitute for your own evals.
  • Positioning: a fast, cost-efficient model for everyday coding and agentic workflows, built for rapid execution across diverse agentic workflows. Per Google's release notes, it adds an xhigh thinking level and is now the default model for the Antigravity agent.

Hydrogen is also, in effect, the successor to the Flash tier: it replaces Gemini 3.8 Flash as the model AI Mode has used, and it is the model Google is putting in front of every Search user. That is the distribution story of this wave. A model can be the best in the world and still not matter if nobody can call it; Hydrogen matters because everyone already can.

The full wave, in one table

Prices below are standard per-million-token API rates (input / output), read from the vendors' own pricing pages on November 13, 2026, as compiled by BeevR — which verified each entry against the vendor's changelog, model page, pricing page, or system card, and deliberately left out rumoured models and aggregator claims.

Vendor Model Date Context Price in / out Who can use it
Anthropic Claude Fable 5.2 9 Nov 1M in, 128K out $8 / $40 All API customers; also AWS, Google Cloud, Microsoft Foundry
Anthropic Claude Mythos 5.2 9 Nov 1M in, 128K out $8 / $40 Verified orgs only (Cyber / Life Sciences programs)
OpenAI GPT-6-Cyber 5 Nov 1.05M (922K max input), 128K out $20 / $100 Approved Daybreak Red organizations
Google Gemini 4 Carbon GA 10 Nov 1M in $2 / $10 until 30 Nov, then $4 / $20 Paid Gemini API, Gemini for business, AI Ultra
Google Gemini 4 Hydrogen GA 12 Nov 1M in $1.50 / $7.50 Free and paid Gemini API tiers

Two things jump out. First, the price spread is enormous: Hydrogen's output token costs less than a tenth of GPT-6-Cyber's. Second, access is the real filter. Of the five models, only Fable 5.2 and Hydrogen are open to all API customers — and Fable costs more than five times Hydrogen per output token.

The planner/worker split

Step back from the individual launches and a pattern appears. PYMNTS, covering the wave on November 12, framed it as Anthropic and Meta splitting the frontier into planners and workers: Anthropic released Claude Fable 5.2 for demanding reasoning and long-horizon agentic work, while Meta released Muse Ember — its smallest and cheapest model — at $0.10 per million input tokens for the repetitive tasks agents do all day.

Fable 5.2 topped the Artificial Analysis Intelligence Index at 65, per PYMNTS, moving past Google's Gemini 4 Carbon. Anthropic's own launch page reports 69.9% on FrontierSWE v2 (against 62.3% for Opus 5.5 and 65.5% for GPT-6 Astra) — again, vendor-run numbers at each vendor's chosen effort level. Fable 5.2 costs 20% less than Fable 5.1 ($8/$40 versus $10/$50).

And yet Anthropic itself says to start with Claude Opus 5.5 and move to Fable 5.2 only when evals at higher effort still fall short. The most capable model is not the default recommendation even from its own vendor. That tells you something about where the frontier actually sits: capability is no longer the bottleneck for most work. Cost, latency, and reliability are.

The catches nobody puts in the headline

Every launch post reads like a gift. The fine print is where the engineering work lives. All of the following come from vendor docs and release notes as compiled in the BeevR roundup:

Fable 5.2 breaks history edits. If your code changes the system prompt, tools, or an earlier message mid-conversation, Fable 5.2 returns a 400 error ("The block is bound to a different conversation"). Agent frameworks that trim or rewrite history need testing before you switch. This is the kind of change that passes every benchmark and breaks a production agent on day one.

Model routers lose reasoning. Only Fable 5.2 and Mythos 5.2 can read Fable 5.2 thinking blocks. A router that moves a conversation to Opus 5.5 drops that reasoning. If your architecture routes between models mid-conversation, that is a real constraint.

No zero data retention by default. Fable 5.2 and Mythos 5.2 require 30-day data retention unless Anthropic expressly authorizes otherwise. For teams handling health or financial data under zero-retention terms, that alone can rule the model out — before benchmarks ever enter the conversation.

GPT-6-Cyber is not for you (probably). It is a cybersecurity variant of GPT-6 Astra for authorized vulnerability research, exploit development, and fix verification. OpenAI rates it Critical in cybersecurity under its Preparedness Framework, reports 48.45% on ExploitGym against 42.43% for GPT-6 Astra, and restricts it to approved Daybreak Red organizations. It is also the most expensive model on this list at $20/$100 — and it does not support the Batch API. In testing, it reportedly found 81 previously unknown high-severity Firefox flaws, per PYMNTS. Impressive, and irrelevant to your next sprint unless you are in the Daybreak Red program.

Carbon's price doubles on December 1. The $2/$10 introductory pricing ends November 30; from December 1 it is $4/$20. Gemini 4 Argon doubles on the same date. If you are load-testing Carbon this month, budget on the December price, not the November one.

Mythos is Fable with fewer safeguards. Fable 5.2 and Mythos 5.2 are the same underlying model with different safeguard settings. Mythos is limited to vetted organizations through Anthropic's Cyber and Life Sciences Verification Programs. The useful detail for everyone else: Fable 5.2's classifiers block about 55% fewer requests than Fable 5.1's did, which cuts false refusals on legitimate security and biomedical work.

The December deadlines matter more than the launches

Several deadlines land within the next four weeks, and for a production system they matter more than any new launch:

Date What happens
30 Nov claude-sonnet-4-5-20250929 retires on the Claude API (move to Sonnet 5.5). Fast mode for Opus 4.8 is removed. OpenAI shuts down its Evals platform, Agent Builder, and reusable prompts (v1/prompts).
1 Dec Gemini 4 Carbon and Argon prices double. gemini-2.5-pro routes to Gemini 3.8 Flash and gemini-2.5-flash to Gemini 3.5 Flash-Lite. Older GPT Image models (gpt-image-1.5, gpt-image-1-mini) are removed.
11 Dec OpenAI removes the gpt-5-2025-08-07 family and o3 snapshots.
1 Jan 2027 Gemini 3.8 Flash and 3.6 Flash prices double.

The routing changes deserve special attention: they are silent upgrades. Your requests keep working, but on a different model. If you have anything pinned to gemini-2.5-pro or gemini-2.5-flash, check what it routes to after December 1 and run your evals against the new target before the switch happens to you.

So should you switch models?

For most teams, no — not on the strength of a launch post. BeevR's roundup puts it well: if you are on Claude Opus 5.5, GPT-6.1 Sol, or Gemini 4 Argon, the November releases do not justify a switch without your own evals. The exceptions are narrow: try Fable 5.2 for long-horizon agent work that already fails at high effort, and re-price anything on Gemini 4 before December 1.

The decision framework they use with clients is worth adopting in full:

  1. Run your own eval set first. 200 to 500 real cases from production logs beat any vendor benchmark.
  2. Price the task, not the token. A model that costs twice as much per token but finishes in half the turns can be cheaper. Count turns, retries, and tool calls — not just the price table.
  3. Check data terms before benchmarks. Retention rules, regional processing, and compliance coverage decide more healthcare and fintech choices than a 3-point benchmark gap.
  4. Keep the model behind an interface. If changing models takes more than a config change plus an eval run, that is the first thing to fix. This month proves the point: prices and lineups change every few weeks.

My take

Here is my honest read, offered as opinion rather than fact: the most important model of November is the cheapest one. Hydrogen at $1.50/$7.50, on a free tier, in front of every Search user, does more to change what developers build than a gated $20/$100 cyber model ever will — because the developers who can call it number in the millions, and the agents they build will be priced by Hydrogen's economics, not Fable's.

The frontier is splitting into two markets. One sells capability to the teams that need the best reasoning on Earth and can pay for it. The other sells good-enough intelligence at a price where you stop counting tokens and start building. Both markets are real. But only one of them is where most of us live.

The other takeaway is less exciting and more useful: the model you run matters less than the layer you run it behind. In a month where five models shipped, three prices changed, and four deadlines landed, the teams that did fine were the ones for whom a model switch is a config change and an eval run — not a rewrite. If that is not your codebase yet, November just gave you the business case.

What do you think?

Which model from this wave are you actually using — and did you switch because of a benchmark, a price, or a deadline? Has a silent routing change or a price doubling ever caught you by surprise in production? And where do you stand on the planner/worker split: one strong model for everything, or cheap models for the routine work and a frontier model for the hard parts?

Sources


Thanks for reading! I'm Shaswat Raj — I build CampusLoop, IndexFast, SopKit and open-source dev tools.
GitHub: https://github.com/playlikeboy/ · LinkedIn: https://www.linkedin.com/in/sh20raj/ · YouTube: https://youtube.com/@stackshade

Top comments (0)