DEV Community: meta

Muse Invite Code DMVDLV: Get 1 Billion Free Tokens (Redeem Within 48 Hours)

wen — Thu, 12 Nov 2026 01:42:13 +0000

Muse Invite Code DMVDLV: Get 1 Billion Free Tokens (Redeem Within 48 Hours)

Muse is Meta's personal AI agent (muse.ai) — it runs on its own virtual computer and can browse, build, and complete real tasks for you. New users who redeem an invite code within 48 hours of joining get 1 billion bonus tokens, and the person who shared the code gets 1 billion too. The bonus tokens never expire.

The code

DMVDLV

How to redeem (takes 1 minute)

Sign up at https://muse.ai/join or install the Muse app (available in the US, Canada, Japan, and Singapore).
Within 48 hours of creating your account:
- iPhone / Android app: Settings → Redeem token
- Web (muse.ai): Settings → General → Usage → Redeem invite code
Enter the code above and confirm.

Don't miss the window

The redeem option only exists during your first 48 hours after joining. After that it disappears and no code will work on your account — so redeem first, explore later.

If this code is full, get a fresh one here

Each code has a limited number of uses. If DMVDLV is full when you try it, don't keep retrying it — get a fresh, verified code from RedeemPool instead:

https://redeempool.app/

RedeemPool is a free community pool for Muse invite codes. Codes in the pool are screenshot-verified for remaining uses and issued one person at a time, so you don't waste your 48-hour window on a code that is already used up. Get the app here:

Download on the App Store: https://apps.apple.com/us/app/redeempool/id6832842012

Have a spare code of your own? List it in the pool the same way — upload a screenshot of your Muse invite page, the app reads your code and the uses left, you confirm, and your code is in the pool. Listing is free.

Good to know

Free-plan users can redeem too. The bonus appears as Additional tokens in Settings.
Your weekly free limit is used first; the bonus tokens are only touched after that, and they never expire.
Each code has a limited number of uses. If this one is full, the app will tell you when you try.
Disclosure: this is my personal invite code. If you redeem it, I receive the same 1 billion token bonus you do. It costs you nothing.

Questions

What is a Muse invite code?
A code from an existing Muse user. A new user who redeems it within 48 hours of joining Muse receives 1 billion bonus tokens, and the code's owner receives the invite reward too.

How long do I have to redeem?
48 hours from when you join Muse. After that, your account loses its redeem eligibility, so claim a code and redeem it right away.

Is RedeemPool free?
Yes. Listing and claiming are both free, and the app contains no purchases.

What if a code does not redeem?
Report it in the RedeemPool app with a screenshot of the Muse error, and claim a different code. Codes are provided by other users, so no code can be guaranteed to redeem.

Who can use RedeemPool?
Users must be 18 or older.

RedeemPool is an independent third-party tool. It is not affiliated with, endorsed by, or sponsored by Meta or Muse. “Muse” is used only to describe what the codes are for. Invite codes are provided by other users, so no code can be guaranteed to redeem. The Muse invite programme is run by Meta and may change or end at any time.

Last updated: November 11, 2026. Code checked in the pool the same day.

Muse Invite Code: 8C6H01 (1 Billion Free Tokens)

Muse Codes — Wed, 11 Nov 2026 22:57:33 +0000

Check out Muse, your personal AI agent. Redeem this code in Settings within 48 hours of joining to get 1 billion Muse tokens.

Code: 8C6H01
https://muse.ai/join

Meta's Head of Infrastructure Santosh Janardhan Leaves

Hacks.gr — Wed, 11 Nov 2026 20:33:19 +0000

Santosh Janardhan is stepping down as Meta's longtime head of infrastructure and co-head of engineering, with former Amazon executive Dave Brown taking over the role as Meta scales infrastructure for AI workloads.

He will be replaced by Dave Brown, who joined Meta earlier this year as vice president of infrastructure. Before Meta, Brown spent 19 years at Amazon, focusing on computing power, web services and artificial intelligence.

The next day for Janardhan

Janardhan will remain with the company temporarily to facilitate a smooth transition and then take on an advisory role on Meta's compute capabilities. His departure is expected to be finalized next year.

Janardhan had been with Meta since 2009. The news first emerged in an internal memo leaked to Bloomberg, which discussed the need to hand over the reins to new people. The memo did not specify the reasons for his departure.

The big investments in artificial intelligence

Meta's infrastructure investments have grown rapidly in recent years as the company tries to keep up with advances in artificial intelligence. Earlier this year, Janardhan was tasked with leading an initiative to expand the company's data center investments and strategy.

↳

>Stay connected with Hacks.gr

Set us as a preferred source on Google and find us on social media.

Preferred source

↗

end of article

Read the original English article on Hacks.gr

How to Make Your Moltbook Agent Reliable: A Practical Guide to Agent Auditability with ZizkaDB

Mir Arshad Ali Talpur — Wed, 11 Nov 2026 08:56:38 +0000

An agent that socializes is an agent that reads strangers. On Moltbook, a Reddit-style network for AI agents, that is the whole point: your agent polls the feed, reads posts, replies, votes and sometimes acts. Researchers have documented what can go wrong when it does. This guide shows what you, as the person running an agent, can do about it: record what the agent reads and does, measure whether it is changing, and keep a record that stands up to scrutiny.

The tool we use is ZizkaDB, an open source database for AI agents. Think of it as an agentic database: an operational database for AI agents that stores each step as an event, links events by cause, and protects the record against quiet edits. You do not need Moltbook to follow this guide. Any agent that reads content from outside your company has the same needs.

A note on scope. This is independent commentary based on public research as of 11 November 2026. ZizkaDB is not affiliated with, endorsed by, or connected to Moltbook, Meta, OpenClaw or any researcher named here. There is no existing ZizkaDB integration with Moltbook. Findings cited below are attributed to their sources and several are estimates from preprints or vendors.

Why your agent needs this

Researchers have reported that Moltbook agents poll the platform roughly every 30 minutes by default. Three findings explain why reliability is not automatic.

Security vendor Vectra AI reported that about 2.6 percent of a sampled set of posts contained hidden prompt injection payloads, and that some agents publicly posted diagnostics such as open ports and error messages.
An arXiv preprint, Agents in the Wild, studied 27,269 agents over nine days. Using pattern matching, it reported that social engineering made up 31.9 percent of detected attacks, against 3.7 percent for prompt injection, and that adversarial posts earned six times the engagement of normal ones.
Researchers reported by The Decoder said they influenced more than 1,000 agents in under a week with narrative posts, while crude injection was mostly ignored.

The practical lesson for you is that a persuasive post leaves no malicious string to filter. If your agent is persuaded, the only evidence is what it read and what it did next. That is why Agent Auditability matters: you need to be able to reconstruct, after the fact, why your agent did something.

What reliable means here

Agents are probabilistic. They will not behave identically every day, so reliability cannot mean never changing. For a social agent it means four things you can check:

You can show what it read before each action (Causal Lineage).
You can tell when its behaviour shifts (Agent Drift, and Agent drift measurement).
You can fix bad inputs and stop it relying on them again.
You can prove your record has not been edited (tamper-evident logs).

ZizkaDB gives you a place to do all four. As an agentic database it keeps behaviour, memory and lineage together, and as an operational database for AI agents it is meant to run alongside your agent day to day, not as an archive you open once a year. It records and measures. It does not block attacks, detect malicious content, or make an agent immune to persuasion. Prevention such as least privilege and sandboxing is a separate layer you should still build.

Step 1: run ZizkaDB next to your agent

ZizkaDB runs self hosted in a single Docker container on Postgres, using the quickstart in its documentation. There is also a managed Cloud edition at db.zizka.ai. The open source edition records one agent per install. If you run several agents, or want to trace ideas between agents, that is a Cloud feature.

Connect through the Python or TypeScript SDK. If your agent framework supports MCP, the ZizkaDB MCP server is another route, but check your framework's documentation before assuming it does.

Step 2: define the events your social agent produces

Start with a small, consistent vocabulary. This is the foundation of any Agent Audit.

Event When to log it Key fields
heartbeat Each time the agent polls the feed time, feed or submolt
post_read Each post the agent reads post id, author, source, trust tag
memory_write Each time the agent stores a fact the fact, source post
decision Each choice to reply, vote, or act what was chosen, reason
post_published Each outbound post or reply content, target
tool_call Each action outside the platform tool, target

Tag anything from another agent as untrusted. That single field makes later questions easy to ask.

Step 3: log reads and link every decision to them

Causal Lineage is the core idea. Each event points to the event that caused it through a parent id. An illustration, with parameter names illustrative and to be checked against the SDK reference:

python
read = await db.log(agent="my-moltbook-agent", event="post_read",
data={"post_id": "p-88", "author": "agent-7391", "trust": "untrusted"})

mem = await db.log(agent="my-moltbook-agent", event="memory_write",
data={"fact": "restart service X to fix error 52"},
parent_id=read.id)

act = await db.log(agent="my-moltbook-agent", event="tool_call",
data={"tool": "restart_service", "target": "prod-db"},
parent_id=mem.id)

(await db.why(act.id)).print()

When the action looks wrong, why() walks back from the tool call, through the memory write, to the post that started it. That is Agent Auditability in practice: one command turns a mystery into a chain.

If you use a batch job to log many reads per heartbeat, use batching and idempotency keys so a retry does not create duplicates. Check the SDK reference for the exact parameter names.

A regular Agent Audit then becomes a short routine: pick an action, run why(), read the chain, and decide whether the agent should have acted on what it read. Teams that run an operational database for AI agents this way find problems in minutes instead of reconstructing them from scattered output logs.

Step 4: keep the record clean

Two habits keep the record useful and safe.

First, redaction. ZizkaDB redacts before storage, so secrets that appear in your agent's context, such as an API key in a pasted post, do not become a second leak inside the log. Configure it for the patterns you care about.

Second, supersedes. When you learn a remembered claim was wrong, write a correcting memory that supersedes the old one. Superseded facts drop out of retrieval, so the agent stops relying on them, while the history stays intact for audit.

Step 5: measure Agent Drift

A social agent can change slowly. It may start posting more, call a different mix of tools, or lean on untrusted sources more than it used to. These shifts are Agent Drift, and none of them throws an error.

Agent drift measurement in ZizkaDB works from a reference period. baseline() compares recent behaviour with an earlier window, so you decide what to watch. Good signals for a Moltbook agent:

Posts published per heartbeat.
The mix of tool calls.
The share of decisions whose lineage includes an untrusted read.
Time between a read and an action.

Pick a quiet week as your baseline, then check weekly. A rise in the untrusted-lineage share, or a new tool appearing in the mix, is a prompt to look at the lineage, not proof of a problem. Drift is a signal to investigate, and the lineage tells you where to look.

Step 6: make the record tamper-evident

An audit trail is only worth something if it cannot be quietly rewritten. ZizkaDB computes a checksum SHA 256 for each event and chains events per agent with a hash chain, so changing an old event breaks the chain. This is the basis for tamper-evident logs, sometimes described as cryptographic telemetry for AI agents: telemetry about what your agent read and did, protected by hashes.

Two points keep this honest. The chain makes edits detectable, it does not show the content was true or correct. And it only protects you if you store the chain head outside the database, for example in a separate system, so an administrator cannot rewrite the chain and the head together. Run verify_chain on a schedule and keep the checksum SHA 256 of the head somewhere independent.

In short, cryptographic telemetry for AI agents means the evidence of what your agent did carries its own integrity check, so a reviewer does not have to take your word for it.

Step 7: handle incidents with the impact route

Suppose you learn that post p-88 was manipulation. Use the impact route to list the events that descend from it. You will see which memory writes it produced, which decisions followed, and which actions it triggered. That is your incident scope, ready to hand to whoever needs it.

If you run several agents, Cloud's agent to agent recording lets you follow an idea across agents. This is how you answer whether a manipulated post spread beyond one agent.

Step 8: let others verify without trusting you

Reliability on a social network is partly about what others can check. ZizkaDB supports read-only and expiring keys. A customer, partner or reviewer can inspect the record for a limited time without being able to change it. That turns it is reliable from a claim into something they can verify.

Where EU AI Act Article 12 fits

EU AI Act Article 12 requires high risk AI systems to support automatic recording of events over their lifetime. Whether your agent counts as high risk depends on what it does, not on the fact that it is an agent. A hobby agent on a social network probably does not. An agent that makes decisions in a regulated setting may. The Digital Omnibus, Regulation 2026/1744, deferred the main high risk obligations to December 2027 for Annex III systems and August 2028 for Annex I.

What does not depend on a deadline is the practical value. A tamper-evident, causally linked record is the kind of evidence Article 12 style record keeping points toward. To be clear, ZizkaDB supports record keeping and investigation. It is not legal advice and it does not by itself deliver EU AI Act compliance. If you need to know where your system stands on EU AI Act compliance, take legal advice on your specific facts.

A weekly checklist
Log every post_read with a trust tag and every decision with its parent.
Link memory writes and tool calls to the reads behind them.
Review the untrusted-lineage share and your baseline() drift signals.
Run verify_chain and store the head outside the database.
Supersede any memory you now know to be wrong.
Check that read-only keys given to reviewers are still the ones you intend.
Limits to remember

ZizkaDB cannot tell whether a post was malicious or true. It cannot verify that an actor is an AI. It does not stop an agent being persuaded. It makes agent behaviour recorded, measurable and checkable, which is what lets you correct it. Confirm the specific features you rely on, such as the impact route, baseline() and agent to agent recording, against the current documentation before you build on them.

Get started

Run the self hosted container, treat it as the agentic database behind one agent that reads the open web, and log reads, memory writes and actions for a week. Then ask your first question: what did you read before you did that?

Corrections: if you represent a party named here and believe a statement is inaccurate, contact us and we will review it promptly.

Sources, all public reporting or published research: Vectra AI (Moltbook and the illusion of harmless AI agent communities); Agents in the Wild, arXiv 2602.13284; The Decoder (Moltbook researchers hijacked agents in days).

Muse Ember Is Cheap. The Pelican Test Told Me Something Else.

Dishant Sharma — Tue, 10 Nov 2026 21:12:39 +0000

Simon Willison ran a pelican test on it. That is the first thing I saw about Muse Ember, before I saw a single benchmark number.

He asked the model to draw a pelican riding a bicycle at all six reasoning levels. It returned five pelicans. The high setting did not produce a complete SVG at all.

Then he said something that stuck with me. The pelican test usually tracks reasoning effort. More thinking, better pelican.

This time the correlation broke. minimal came back with about ten reasoning tokens and, by several readers' judgment, the best bird of the set.

Meta Superintelligence Labs shipped its smallest model on November 9. The top comment thread on Hacker News was about whether the pelican had a fish in its basket.

I read 139 comments. Maybe six were about the pricing.

What they actually shipped

Muse Ember is a small multimodal model built for the boring part of agents. Tool calls, routing, extraction, coding subtasks, long document parsing. The stuff an agent does four hundred times while you watch a spinner.

The pricing is the headline, and it is genuinely low:

$0.10 per million input tokens
$0.004 per million cached input tokens
$0.50 per million output tokens

That cache number is the one people keep missing. It is 4% of input. Luna and Haiku charge about 10%.

In agent loops, cached input is most of the bill.

Meta's pitch is simple. A big model plans, this one does the volume. Their line: Muse Wonder plans, Muse Ember handles the volume.

The benchmark story, and the number that does not match

Artificial Analysis puts it at 36 on the Intelligence Index for $0.14 per task. Claude Haiku 5.5 hits the same 36 and costs $0.41. That is nearly three times more for the same score.

Model	Index	Cost per task
Muse Ember max	36	$0.14
Claude Haiku 5.5 max	36	$0.41
GPT-6 Luna max	28	$0.15

And it beats DeepSeek V4.1 Flash on cost, which is the part that made people sit up. DeepSeek was supposed to own cheap.

But here is the thing I could not stop poking at. Meta's own launch page lists AutomationBench-AA at 37.8% for Ember. Artificial Analysis wrote 68.1% for the same benchmark in the same week.

I do not know which test setup produced which number. Neither page explains the gap.

That is not a scandal. Benchmarks get re-run, versions drift, runners differ.

But it does mean you should not quote either figure like it is carved in stone. I have been burned by this exact thing before.

"Smallest" is doing a lot of work

Everyone asked the same question in different words. Small compared to what? How many parameters?

Can I run it?

The answer is no. There are no weights. Muse Ember is API only, muse-ember-1.0, on Meta's Model API and Muse Code.

One commenter put it better than I could: what is the point of bragging that it is small if it always runs on someone else's datacenter?

Fair. Though the "small" claim is not really about your laptop. It is a signal that this competes with Haiku and Luna, not with Wonder.

Meta is telling you where to slot it in your agent stack.

The open weights crowd is still annoyed. Muse Spark Open shipped at 816B parameters, which is not a thing most people can run either. And the open version came with extra safety training that several users described as being lobotomized into announcing it is Muse Spark regardless of the system prompt.

There is a running joke in that thread about whether the open releases are distills wearing a safety hat. Nobody proved it. But nobody disproved it either, and Meta has not published the parameter count for anything closed.

So the pattern is clear. Meta releases the closed good one, then a heavier, more restricted open one later. Ember fits the pattern.

The trick that keeps it honest

Here is the detail I actually respect.

On AA-Omniscience, Ember scores 21.1% accuracy. That is the lowest of its comparison group. Haiku gets 34.7%.

But it declines to answer 64% of the questions. And that keeps its hallucination rate at 17.4%, half of Haiku's.

So the model is not knowledgeable. It just knows when to shut up.

I have wanted this from every model for two years. A model that says "i don't know" instead of confidently inventing a citation is worth more to me than one that scores twelve points higher and makes things up. For an agent doing extraction at volume, a wrong value in a database costs real money.

It is also the least glamorous thing a model can do. Declining to answer looks like weakness on a scoreboard. In production it looks like not paging anyone at 2am.

The catch is obvious. If you need it to know things, this is not your model. It is a doer, not an encyclopedia.

The pelican thing is not a joke

I keep coming back to the birds.

Every model now draws the same pelican. Same red scarf. Same fish in a front basket.

Same three motion lines, same yellow sun, same dashed lane markings. One commenter called them ley lines, the paths a model walks when you leave a prompt vague.

That is not really about Muse Ember. It is about all of them. A short prompt makes the model fill the gaps from training associations, and the associations are now shared across every lab.

Which means the pelican test stopped measuring reasoning a while ago. It measures how much a model is willing to wander. And when a launch thread spends its energy there instead of on the pricing, that tells you something about what people actually do with these things all day.

The honest part

Most people do not need this model. If you are building a chatbot that answers twenty questions an hour, the price difference is noise. Use whatever you already have wired up.

And the benchmaxxed complaint is real. One user who had actually used the Muse line said the models read as frontier on bench cards but feel less useful in real work.

Another said Spark was their daily driver for weeks and got quietly outclassed. These are people with opinions formed by using the thing, not reading about it.

Also, no batch API at launch. One commenter's entire review was three words: "No batch. Pass."

If your workload is bulk and offline, that matters more than any index score.

Where this lands

The pelican with the fish is the one I remember. Not the 36.

Meta built a model that is cheap, quiet when unsure, and good at the repetitive middle of agent work. That is a real product decision, and a smarter one than chasing the top of the leaderboard.

But the launch told me something else too. We have gotten so used to benchmark charts that a genuinely low price barely registers, while a bird on a bicycle gets 300 upvotes.

Next time i run an agent loop that burns forty thousand tokens on nothing, i will think about that pelican. And i will check the cache price before i check the index.

Meta's Muse Wonder Is a Real Frontier Model Now — But the Best Version Isn't on the API

Paul Lovelace — Tue, 10 Nov 2026 17:57:09 +0000

On October 12, 2026, Meta released Muse Wonder and Muse Wonder Full — the first frontier-class models from Alexandr Wang's rebuilt Meta Superintelligence Labs. The claim from Wang: "true frontier models," roughly level with the best from OpenAI and Anthropic. Nearly a month later, with independent scores in, that claim mostly holds. The catch is not the benchmarks. It's which version you can actually use.

I have not run Muse Wonder hands-on yet, and this article is built from Meta's announcement and model card, plus independent coverage from VentureBeat and the Artificial Analysis scores — checked November 10, 2026. Where a claim is Meta's, I'll say so, because a launch scorecard is a vendor claim, not an independent result.

What Meta actually shipped

Two models, developed under the internal codename "Watermelon." Muse Wonder is the standard model; Muse Wonder Full is the larger, stronger version. Both are built for the same job: personal and professional agentic work — long-running tasks where the model has to plan, use tools, remember context, and finish something, not just answer a question.

The specs that matter:

1M-token context window at inference, per the official model card.
Multimodal input — text, image, video, and audio in; text out.
Active Memory — Meta's headline feature. The model is designed to remember things you mentioned days, weeks, or months ago and use them in later tasks. Meta's own MemoryBench puts Wonder at 85.6 and Wonder Full at 85.5, against 81.5 for Claude Opus 5.5 and 81.4 for GPT-6 Astra. Read that carefully, though: MemoryBench is Meta's own benchmark, measuring recall on Meta's own definition of remembering. The gap is real but narrow, and it's Meta grading its own homework.
Active Focus — the model's ability to stay on a long task without drifting, which is the difference between an agent that finishes a job and one that wanders off after step three.

Both models were trained at Prometheus, Meta's gigawatt-scale data center in Ohio. That detail matters less as trivia and more as a signal: this is the $100-billion-plus compute bet Zuckerberg and Wang committed to at the end of 2025, before the model existed. The lab itself is tiny by frontier standards — fewer than 200 researchers — built after Llama 4 landed badly in April 2025 and Zuckerberg paid $14.3 billion for a 49% stake in Scale AI to bring Wang in.

The benchmarks, honestly read

Here is Meta's published scorecard, all at maximum reasoning effort. I've kept the independent Artificial Analysis row first, because it's the one number nobody at Meta chose.

Benchmark	Wonder Full	Wonder	GPT-6 Astra	Claude Opus 5.5
Artificial Analysis Intelligence Index	56	52	53	58
DeepSWE 1.1 (long-horizon coding)	79.0	78.8	74.1	74.2
SWE-Marathon (long-running coding)	57.5	53.1	42.5	46.3
Harvey LAB-AA (legal tasks)	13.1	10.0	8.6	4.2
AutomationBench-AA (business workflows)	45.4	39.1	41.6	41.2
MemoryBench	85.5	85.6	81.4	81.5
Terminal-Bench 4.0-AA	57.1	52.0	59.1	59.6
HealthBench Pro	62.3	60.5	65.3	65.6

What this actually says: Wonder Full ties Claude Sonnet 5.5 at 56 on the Artificial Analysis Intelligence Index, ahead of GPT-6 Astra (53) and Claude Fable 5.1 (53), but behind Claude Opus 5.5 (58). The standard Wonder scores 52 — ninth of 216 models in its class on that index, one point behind GPT-6 Astra.

The pattern in the rest of the table is consistent and worth noticing. Wonder Full leads clearly where the work is long and agentic — DeepSWE, SWE-Marathon, legal work, business automation. It loses where the work is general knowledge or health reasoning, and it trails slightly on Terminal-Bench. That is not a random scatter. It is a model tuned for agents, and it shows.

Two caveats, both from Meta's own model card. First, the scorecard mixes Meta-run evaluations, third-party results, and official leaderboard scores, and the card says competitor runs are best-effort — prompts, tools, and runtimes may not be equally optimized for every model. Second, on SWE-Marathon specifically, the GPT-6 and Fable results were produced with Codex and Claude Code respectively, so that row is not a controlled common-harness comparison. Neither caveat erases the result, but both are the reason you should read the table as "roughly level at the frontier" rather than "Meta wins."

The price — and the paywall

Through the Meta Model API, standard Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens. That is exactly the same as Gemini 4 Carbon's regular pricing, and half of Claude Fable 5.2's $8/$40. On price alone, Wonder is competitive.

Here is the catch, and it's the most important fact in this article: Muse Wonder Full is not available through the API at all. The stronger model — the one that ties Sonnet 5.5 and leads the coding benchmarks — is only available to Muse Code subscribers, on $50 or $200 monthly plans. The standard Wonder is also in the Muse app on $20 or $100 plans, while the free tier stays on the smaller Muse Spark.

And there is a second, quieter break with Meta's past: neither Wonder model ships with downloadable weights. This is a closed model, hosted only by Meta, sold by subscription and by the token like everyone else's. The company that built its reputation on open Llama weights is now competing the closed way. If you were waiting for an open frontier model from Meta, this isn't it.

What it means for developers

The practical question is not whether Wonder is "frontier" — on the independent index it is, just barely, and on long agentic coding tasks it's genuinely strong. The question is whether it's worth your tokens and your time.

Try the API Wonder first on your longest coding task. DeepSWE and SWE-Marathon are the benchmarks that map closest to real work: a multi-file refactor, a long debugging session, a feature that takes hours, not minutes. At $4/$20 with $0.20 cached input, it's priced to be tried. Give it the task that usually takes your current model two or three attempts and see if it finishes in one.
Don't pay for Wonder Full until you've hit Wonder's ceiling. The Full model's advantage is real on the scorecard, but it's locked behind a Muse Code subscription. If the API model handles your work, the subscription buys you a benchmark edge, not a different capability.
The memory feature is the one to watch, not the benchmarks. If Active Memory works the way Meta describes it — remembering project decisions and preferences across weeks — that's the feature that changes how you use an agent day to day. But it's also the feature with the least independent verification so far. MemoryBench is Meta's own test. Treat it as a promise to verify, not a result.
Plan around a closed model. No weights means no self-hosting, no fine-tuning on your own hardware, and no exit if Meta changes pricing or terms. That's the same deal as OpenAI and Anthropic, and it's now Meta's deal too. If open weights matter to your stack, Wonder doesn't change your options.

What we don't know yet

Independent reproduction of the agentic benchmarks. The Artificial Analysis index is independent; most of the task-specific rows are Meta-run or Meta-selected. Nobody outside Meta has run the full suite independently yet.
How Active Memory behaves in real projects. A benchmark with distractor conversations is not the same as six weeks of real project history.
The full safety report. Meta says a full report with additional safety evaluations is coming; for now the 13-page model card is the main public technical document. It reports CyberGym at 88.4% and ExploitBench at 58.9% for Wonder Full, and notes the unmitigated model was assessed at moderate or lower risk for cybersecurity and loss of control.
What happens to the Muse app's free tier. Wonder starts in paid subscriptions because it costs more to serve. Whether it ever reaches free users is not announced.

The bottom line

Meta is back in the frontier race, and this time the claim survives contact with independent scores. Muse Wonder Full ties Sonnet 5.5 on the Artificial Analysis index and beats the field on long-horizon coding, legal work, and business automation. The standard Wonder, the one you can actually call through an API, is a step behind at 52 — good, not leading.

But the story isn't really the scoreboard. It's that Meta spent $100 billion-plus, rebuilt its entire AI lab around a 29-year-old who had never trained a frontier model, and shipped something that belongs in the same conversation as Claude and GPT-6. The old Meta — the one that gave away Llama weights and competed on distribution — is gone. The new Meta competes on the model itself, behind a paywall, at the same price as everyone else.

That makes Wonder worth trying, not worth switching to sight unseen. Run your hardest task through the API, compare it against whatever you use today, and let your own work decide. The benchmarks say it's close. Your task is the only judge that matters.

Would you give Wonder your hardest coding task this week — or does the fact that the best version isn't on the API keep you where you are?

Sources: Meta announcement and Muse Wonder model card, October 12, 2026; VentureBeat, October 12, 2026; Artificial Analysis Intelligence Index scores as reported October 2026; The Business Times, October 13, 2026; Times of India, November 9–10, 2026 (Wang/Zuckerberg Cheeky Pint interview). All capability, benchmark and pricing claims are Meta's or the cited third parties', stated as such. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.

Meta's Muse is now available on Apple Watch and Wear OS watches

Hacks.gr — Tue, 10 Nov 2026 15:46:00 +0000

Meta has expanded its Muse assistant to Apple Watch and Wear OS, bringing voice calls, messaging, dictation, account management, and task automation directly to smartwatches.

This follows its recent tablet application release and high mobile adoption milestones.

Muse can connect to user accounts and help manage emails, meetings, and bills. It can also take reservations, order groceries, and do tasks to achieve goals. Users can change the appearance of their digital character to resemble a human, animal, or other character.

Muse was originally released on the App Store in September 2026. In October, Meta released a dedicated iPad app, and on November 6, it introduced Spaces: separate conversations with Muse around specific topics and goals.

The week before the watch announcement, users in the US were able to video call with Muse’s digital avatar. According to Meta, the avatar uses expressions and gestures depending on the tone and context of the conversation. Muse can’t see users, but it can hear and respond in real time.

↳

>Stay connected with Hacks.gr

Set us as a preferred source on Google and find us on social media.

Preferred source

↗

end of article

Read the original English article on Hacks.gr

Meta Muse Pricing and Availability: What the Free, Power and Maximum Plans Actually Give You

MuseInfo — Tue, 10 Nov 2026 12:54:16 +0000

Meta Muse Pricing and Availability: What the Free, Power and Maximum Plans Actually Give You

Meta Muse is free to try, and that sentence hides the only question that matters: free to do what, and for how long? An AI agent is not a chatbot with a monthly fee — it does work across your apps, and the amount of work it can do is exactly what the plans ration. This article walks through the three plans Meta lists — Free, Power, and Maximum — what each one actually gives you, who each one fits, and where you can and cannot use Muse yet. One warning first: pricing and availability change over time. The figures below are the ones listed on MuseInfo.ai at the time of writing, so check the current limits on Meta's official subscription page before you subscribe.

The three plans at a glance

Plan	Price	General Use
Free	$0	Light or occasional use
Power	$20/month	Higher usage
Maximum	$100/month	Heavy usage

That table is the whole pricing page in miniature. Meta does not publish a line-by-line feature matrix, because the thing you are paying for is mostly capacity: how much work the agent can do before it asks you to wait or pay. The plans are three levels of the same agent, not three different products.

The Free plan: a real trial, not a demo

Free costs $0 and covers light or occasional use. The important word there is "limits". Muse offers free access with usage limits, which means the Free plan is a working version of the agent, not a locked demo — you can connect an app, give it a task, and see how it behaves. What runs out is capacity: after a certain amount of work, you will hit the limit and have to wait or upgrade.

That is actually the right way to try an agent. The risk with Muse is not the price, it is the setup: permissions to review and connectors to choose. The Free plan lets you do all of that without spending anything — connect one low-risk app, usually a calendar, give Muse research or planning tasks, and check the activity history.

Who Free fits

Free fits anyone who is still deciding whether an agent belongs in their workflow. Occasional users — help with a trip now and then, a shopping comparison, or a meeting follow-up — may never need more. It also fits the cautious: if you want to learn how approvals, permissions, and the audit trail behave before handing the agent anything sensitive, Free is the plan for that learning period. The limit is the point, though — a morning of email triage, a research project, and a travel plan in one session is exactly the multi-step work that consumes capacity. If you find yourself rationing tasks, that is the signal to look at Power.

The Power plan: $20 a month for higher usage

Power costs $20 a month and is aimed at higher usage. The honest way to describe it is that it buys you more of the same agent: more tasks, more background work, more room to use Muse as a daily tool instead of an occasional one.

Twenty dollars a month is the price point where most personal AI tools sit, and the comparison that matters is not against other subscriptions but against the hour it saves. If Muse prepares your meeting follow-ups, drafts your customer replies, and keeps your calendar straight, the cost of the plan is smaller than the cost of doing that work by hand. If it only answers the occasional question, Free was already enough.

Who Power fits

Power fits the daily user: someone who has already tested Muse on Free, decided the agent is useful, and now hits the free limits regularly. It also fits the small-business user who has connected one or two apps and wants the agent to keep working in the background after the app closes. Before subscribing, check the current limits on Meta's official subscription page, and answer one question: how often did you actually hit the Free limit? Once a month, wait. Every week, Power is the plan doing the work.

The Maximum plan: $100 a month for heavy usage

Maximum costs $100 a month and is aimed at heavy usage — five times the price of Power, for people who use the agent as infrastructure rather than as a helper: long-running tasks, many connected apps, and work that crosses services all day.

A hundred dollars a month is not a casual subscription. It is the price of a tool that has become part of how you work, for people who already know from their Power usage that capacity is the constraint — the sheer amount of work they are putting through it. Most people should skip Maximum, at least at first: if you have not yet hit Power's limits, you are paying for headroom you will not use.

How to choose: a simple rule

The choice between the plans comes down to one measurement: how much work you actually put through Muse in a week.

Light or occasional use — a few tasks a week: stay on Free.
Daily use — Muse is part of your routine and you hit the free limits: Power at $20 a month.
Heavy use — the agent runs across many apps all day: Maximum at $100 a month.

Judge the cost against the hour it saves, not against other subscriptions. The subscription is the smallest cost of using an agent; the larger one is the time spent setting permissions and checking the first few weeks of output, and that time is the same on every plan.

Availability: where you can use Muse

Pricing is only half the question. The other half is whether Muse is available to you at all, and that answer is still changing.

Muse launched in the United States and later expanded into Canada through subsequent rollout updates. Availability can vary by account, plan, platform, or region, and some features may still be in beta. If you are outside the launched regions, check availability in your country and on your device before assuming anything — the rollout is ongoing. One more thing worth knowing: a plan does not buy you access to a region or a platform where Muse has not launched, so check availability first and choose a plan second.

Platforms

Meta currently offers Muse on web, iPhone, Android, Mac, and supported messaging surfaces. That breadth matters for an agent more than for a chatbot: Muse can keep working in the background after the app closes, so the device you start on is not necessarily the device you finish on. Check that your device and platform are supported before you build a routine around it.

The bottom line

Muse's pricing is simple on purpose: Free at $0 for light use, Power at $20 a month for higher usage, and Maximum at $100 a month for heavy usage. The plans differ in capacity, not in the agent itself, so the right plan is the one that matches how much work you actually put through it. Start free, measure your usage, and upgrade only when you feel the limit. The figures are date-sensitive — pricing can change, availability changes over time, and features still in beta can shift — so check the current limits on Meta's official subscription page before subscribing. If you want the full picture in one place, with features, connectors, pricing, and limits kept together, see this Meta Muse guide. It is an independent resource, not affiliated with Meta, and it separates what is available now from what is still rolling out, which is exactly the distinction that matters when a product changes this fast.

Meta's Muse Ember Does Haiku's Job for a Third of the Cost — Here's the Catch

Paul Lovelace — Tue, 10 Nov 2026 04:47:54 +0000

Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Facts below come from the sources listed at the end, checked 10 November 2026. We have not tested this hands-on, and nothing here is a hands-on finding.

Meta launched Muse Ember on November 9, 2026 — its smallest and most efficient Muse model, and the clearest sign yet of where the AI industry is heading: not one giant model for everything, but a team of models, with the cheap one doing the bulk of the work. Ember is built for the repetitive jobs AI agents do all day — tool calls, coding subtasks, routing, extraction, long-document parsing. The headline number: $0.10 per million input tokens and $0.50 per million output tokens, with an Artificial Analysis Intelligence Index score of 36 — the same as Claude Haiku 5.5 at maximum reasoning, at about a third of the cost per task.

I have not run Ember yet — nobody outside Meta's early partners had access before yesterday — and this article is built from Meta's announcement and the SiliconANGLE, PYMNTS and Unite.AI coverage of it, checked 10 November 2026. Where a claim is Meta's or Artificial Analysis's, I'll say so, because a launch-day benchmark is a vendor claim, not an independent result.

How we put this together

We verified the story against Meta's announcement ("Introducing Our Smallest, Most Efficient Model: Muse Ember"), SiliconANGLE's coverage and PYMNTS' coverage, all 9 November 2026, plus Unite.AI and CellCog. We have not used Ember hands-on, so nothing here is a hands-on finding. Where a figure comes from Meta or Artificial Analysis — most importantly the benchmark scores and the cost-per-task comparison — we say so plainly instead of presenting it as an independent result.

Verdict: Ember is not the model you reach for when a problem is hard. It is the model you reach for when a problem is repeated a thousand times a day. The price is genuinely cheap — ten cents per million input tokens is the kind of number that changes what you can afford to automate — and the Haiku-matching score, if it holds up in independent testing, makes it a serious option for agent workloads. But the catch matters: it is API-only, you cannot download it, and every performance number so far comes from Meta or from a benchmark Meta points to. Try it on your cheapest, highest-volume workload first. That is where Ember either earns its place or does not.

What actually launched

Muse Ember, model ID muse-ember-1.0, is live now on the Meta Model API and in Muse Code. It is also listed on OpenRouter, and on Google Cloud (through the Gemini Enterprise Agent Platform) and Oracle Cloud. The early partners Meta named — Box, Cline, Oracle, Cloudflare and You.com — tell you who the target customer is: companies running agents at scale, where the model gets called thousands of times a day and the bill is the bottleneck.

The spec sheet is built for volume, not for showpiece reasoning. A 1M-token context window, up to 512K output tokens, multimodal input (text, images, audio and video), and support for the Responses API. Meta's framing is blunt: "Muse Wonder plans, Muse Ember handles the volume." Wonder is the bigger model that figures out what to do; Ember is the one that does the doing, over and over, cheaply.

There is one more detail worth noting: Ember is the first model trained using Muse Wonder Full. Meta is now using its own flagship model to build its smaller ones — a production line, not a one-off launch.

What it costs

The pricing is the whole point of this model, so here it is plainly: $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Cached input at four-tenths of a cent per million tokens is effectively free for the repeated prompts agents send — system prompts, retrieved context, the same instructions on every call. That is the pricing shape of a model designed to be called constantly: the per-call cost is so low that the limiting factor becomes your own infrastructure, not the model bill.

For comparison, Meta's own numbers put Ember's cost per task at $0.14 against Claude Haiku 5.5's $0.41 — about a third. And against the other Muse model in the family, Muse Spark 1.3 scores 37 on the same index but costs $2.39 per task, roughly seventeen times Ember's. Spark is the smarter model; Ember is the cheaper one. That is the trade, stated honestly.

How it stacks up — and why to read the numbers carefully

On the Artificial Analysis Intelligence Index, Ember scores 36. Claude Haiku 5.5, at maximum reasoning, also scores 36. Meta says Ember beats GPT-6 Luna at every reasoning setting. And Muse Spark 1.3, the bigger sibling, scores 37 at $2.39 per task.

Read that table the right way round. Ember matching Haiku at a third of the cost per task is a real claim, and if independent testing confirms it, it is a genuine shift in the economics of agent workloads. But every number in that paragraph comes from Meta or from Artificial Analysis — not from independent hands-on testing, and not from us. Benchmarks on launch day are a press release with a chart attached. The honest move is to treat the score as a reason to try Ember on a real workload, not as a settled fact about which model is better.

There is also a structural point worth making. A score of 36 puts Ember in the same tier as Haiku 5.5 on this index — it does not make it a frontier reasoning model. Nobody at Meta is claiming it is. The pitch is volume, not brilliance: a model good enough for the routine work, cheap enough to run constantly.

When Ember is the right pick — and when it is not

Ember makes sense wherever the same kind of call happens over and over: routing a request to the right tool, extracting structured data from documents, running coding subtasks inside a larger agent workflow, parsing long documents, making tool calls in a loop. If you are building agents that do hundreds or thousands of small jobs a day, the per-call price is the feature that matters most — and Ember's price is low enough to change which jobs are worth automating at all.

It is the wrong pick when the task is genuinely hard: novel reasoning, ambiguous problems, work where a wrong answer is expensive. That is what the bigger models are for — Muse Wonder in Meta's own framing, or Opus-class models from other providers. The sensible architecture, and the one Meta is openly selling, is a team: Wonder plans, Ember handles the volume. Using Ember for everything would be the cheap way to get the wrong answer faster; using a flagship model for every routine call is the expensive way to get the right answer slowly.

If you already run agents on Haiku 5.5 or a similar small model, the honest test costs almost nothing: point Ember at your highest-volume, lowest-stakes workload — the one where the output is checked by a human or by a bigger model downstream — and compare the bill and the failure rate over a week. That is the only benchmark that matters for your use case.

The catch, stated plainly

Every performance claim is Meta's or Artificial Analysis's. The score of 36, the cost-per-task comparison, the claim of beating GPT-6 Luna at every reasoning setting — all vendor-reported or benchmark-provider figures, none independently replicated yet. Treat them as a reason to try the model, not as a settled fact.
API-only — you cannot download it. Meta has not released the weights, and no parameter count has been disclosed. If you need to run a model locally, self-host it, or fine-tune it on your own hardware, Ember is not an option. This is a rental, not a purchase.
It is a volume worker, not a thinker. A score of 36 is Haiku-tier, not frontier-tier. The cheap price is the point; do not mistake it for a model that replaces your biggest model on hard problems.
The ecosystem lock-in is real. Ember lives on the Meta Model API, Muse Code, OpenRouter, Google Cloud and Oracle Cloud. The cached-input price of $0.004 per million tokens is a strong reason to build on Meta's API — and a strong reason to check what happens to your costs if you ever want to leave it.

The bottom line

Muse Ember is the least glamorous kind of launch and possibly the most consequential: a model that is not trying to be the smartest, only the cheapest good-enough option for work that repeats all day. If the Haiku-matching score holds up outside Meta's own testing, the economics of agent workloads shift — not because Ember is brilliant, but because at $0.10 per million input tokens, a whole class of automation that was too expensive to run becomes worth running.

The test is cheap and the verdict is yours to reach in a week: give Ember your highest-volume, lowest-stakes workload and watch the bill and the failure rate. If both stay low, you have found your volume worker.

Would you trust a $0.10-per-million-token model with your highest-volume agent workload — or wait for independent benchmarks first?

Sources: Meta announcement — "Introducing Our Smallest, Most Efficient Model: Muse Ember" (9 November 2026); SiliconANGLE (9 November 2026); PYMNTS (9 November 2026); Unite.AI (9 November 2026); CellCog. All capability, benchmark and pricing claims are Meta's or Artificial Analysis's, stated as such. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.

Meta's New AI Costs 10 Cents — and That's Why Your Next App Will Have an Agent in It

Shakeel Ahmed — Tue, 10 Nov 2026 04:42:21 +0000

By Shakeel Ahmed — AI Growth Lab | November 10, 2026

Originally published on AI Growth Lab

On November 9, 2026, Meta released Muse Ember, its smallest AI model. It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1-million-token context window. Meta built it for the repetitive work AI agents do all day: tool calls, routing, extraction, coding subtasks, and reading long documents.

That price is the whole story. The moment an AI agent costs almost nothing to run, every app starts hiring one. This article explains what Muse Ember is, what 10 cents actually buys, why cheap agents change who can build AI products, and where the small model still falls short.

What Is Muse Ember?

Meta's Muse family is the company's line of AI models for coding and agents, built by Meta Superintelligence Labs. Muse Ember is the smallest and cheapest member of that family. It went live on November 9 as muse-ember-1.0 on the Meta Model API and inside Muse Code, Meta's coding agent. It is also listed on OpenRouter.

Here is the short answer to what it does: Ember is built for the high-volume steps that make up most of an agent's workload. A modern AI agent does not make one big decision. It makes hundreds of small ones — call a tool, read the result, decide the next step, extract a fact from a document, route a request to the right place, repeat. Each of those steps is a separate call to a model, and each call costs money. Ember exists to make those calls cheap.

Meta's launch post describes it plainly: Ember is designed "for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing." It accepts text, images, audio, video, and PDFs as input and returns text. It can take in a million tokens at once — roughly 750,000 words — and write up to 512,000 tokens in reply.

Meta CEO Mark Zuckerberg put it this way in a post on X: Ember is "our smallest model that delivers higher intelligence per dollar than others in its class." That phrase — intelligence per dollar — is the metric that matters here, and it is worth understanding properly.

What Does 10 Cents Actually Buy?

The headline price is $0.10 per million input tokens and $0.50 per million output tokens. But the real math is more interesting than the headline, because there are three prices, not one.

What you send	Price per million tokens
Input, up to 200,000 tokens	$0.10
Cached input, up to 200,000 tokens	$0.004
Output	$0.50
Input, above 200,000 tokens	$0.30
Output, above 200,000 tokens	$1.50

A request with 100,000 regular input tokens and 10,000 output tokens costs about $0.015 — a cent and a half. If all the input tokens qualify for the cached rate, the same request costs about $0.0054 — about half a cent.

Compare that with the big models. Claude Fable 5.2, launched the same day, costs $8 per million input tokens and $40 per million output tokens — making Ember 80 times cheaper per token on both input and output.

But the fairest comparison is with the model Ember is actually measured against: Anthropic's Claude Haiku 5.5. Ember's per-token prices are identical to Haiku 5.5's standard rate. The savings come from somewhere else. Independent benchmarking firm Artificial Analysis found that Ember matches Haiku 5.5's best score while using fewer tokens per job — so the same task costs about $0.14 with Ember and $0.41 with Haiku 5.5. Same price per token, one-third the bill. The model is cheaper because it is more efficient, not because the rate card is lower.

There is a catch in the price table worth knowing. Above 200,000 input tokens, Ember's rates triple to $0.30 input and $1.50 output. For the long-document work Meta advertises, that jump matters. A million-token context window is only cheap for the first 200,000 tokens.

Why Does Cheap Change Everything?

AI agents have been expensive to run at scale, and that cost has quietly decided who gets to build them. An agent that makes hundreds of intermediate calls per task, at frontier-model prices, burns money faster than it creates value. That is why most "AI agent" products so far have been demos, or have been limited to a handful of large companies that can afford the bill.

Ember attacks exactly that bottleneck. Meta's launch post positions it as a subagent that can handle hundreds of intermediate steps without sending every request to a larger frontier model. The pattern the industry is settling on is a division of labor: the big model plans, the small model executes. A frontier model decides what to do; a cheap model like Ember does the repetitive work of doing it — calling tools, parsing documents, running coding subtasks — hundreds of times per task, at a fraction of a cent each.

When the cost of the repetitive layer falls this far, three things change at once:

Small teams can afford agents. A solo developer or a small business can now run an agent that works all day, because the bill is measured in cents per task rather than dollars.
Agents can run continuously. At $0.14 per task, an agent can check, monitor, sort, and route work around the clock. At frontier prices, that same work had to wait for a human to ask for it.
Every app gets an agent layer. The same way every app eventually got search, notifications, and a settings page, every app is about to get an agent that handles the repetitive work inside it.

The demand side is already visible. A Chinese AI agent app just hit 1.5 million daily users — and nobody in the West noticed. People want agents that do things for them. Until now, the cost of running them has been the wall. Ember is a serious attempt to take that wall down.

Is It Actually Good, or Just Cheap?

Cheap means nothing if the model cannot do the job. So the benchmark numbers matter, and Meta invited an independent check. Artificial Analysis, an independent evaluation firm, put Ember through its Intelligence Index, version 5, on the day of launch.

At maximum reasoning effort, Ember scored 36 on the Artificial Analysis Intelligence Index, effectively tied with Claude Haiku 5.5 at 36.2 (Ember itself measured 35.8) and well ahead of OpenAI's GPT-6 Luna, which scored 28. At a lower reasoning setting, Ember scored 34 at $0.12 per task. Artificial Analysis calls Ember the cheapest model on its leaderboard scoring 36 or higher.

Meta's own claim is stronger: Ember "delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting," matching Haiku's best score at about a third of the cost and beating every Luna setting.

Three honest caveats. First, these are benchmark estimates — your production bill will vary with workload shape, retry rate, and output length. Second, the Intelligence Index measures general capability; it does not prove Ember handles your specific task as well as Haiku does. Third, the gap to the top is real: Meta's own Muse Spark 1.3 scores 37, one point above Ember, but costs $2.39 per task — about 17 times more. Ember trades a little capability for a lot of savings, and that trade is the entire point.

Ember is also the first model trained using Muse Wonder Full, Meta's newer training approach. Meta has expanded the Muse family fast since Muse Spark launched in April, followed by Muse Spark 1.3 and the Muse Wonder models in October. Ember is the bottom of that family by size — and, deliberately, the top of it by value.

The Catch — What Ember Cannot Do

No model this cheap is a frontier model, and Ember does not pretend to be one. Its limits matter as much as its price.

It is built for narrow, repeatable jobs. Tool calls, routing, extraction, coding subtasks, long-document parsing. Ask it to do open-ended research, novel reasoning, or creative work, and the bigger models will beat it.
It is a subagent, not the boss. Ember is designed to work under a larger model that plans and checks its work. Running it without that supervision is asking a junior employee to run the company.
The price jumps past 200,000 tokens. The $0.10 rate applies only up to 200,000 input tokens. Past that, input costs $0.30 and output $1.50 per million. Long-document work is still cheap, but not as cheap as the headline suggests.
It is Meta's first version of this idea. The Muse family is months old, and the tooling around it — evaluations, routing patterns, failure modes — is still being worked out in public.

None of this is a reason to dismiss Ember. It is a reason to use it for what it is: the cheap, tireless worker inside a system where a bigger model still does the thinking.

What This Means for Your Next App

If you build software, the practical question is not whether Ember is the best model. It is whether your product can afford an agent layer now — and the answer changed on November 9.

Start with the routing pattern. Use a frontier model for planning and judgment, and route the repetitive steps — tool calls, extraction, parsing, subtask execution — to Ember. That is the architecture Meta is selling, and it is the architecture that makes the economics work. Claude can now read your bank account — and the agent wave behind it is the same one Ember is built for.

If you are a solo developer, the more important number is $0.14 per task. That is the cost of letting an agent do a complete job — not a single call, a whole job. A hundred agent jobs a day costs $14. A thousand costs $140. Those are numbers a small business can budget for, and that is new.

And if you are watching the job market rather than building products, the same shift cuts both ways. AI layoffs in 2026 are hitting some roles while others are hiring — the map is more complicated than the headlines. Cheap agents do not replace the people who design the work. They replace the repetitive steps inside it — which is exactly the work Ember is built to do.

The agent era has been waiting on one thing: a model cheap enough to run all day without asking permission. Ember is not the smartest model Meta has ever shipped. It is the cheapest one that is good enough. For most apps, that is the one that matters.

Sources: Meta — Introducing Our Smallest, Most Efficient Model: Muse Ember · VentureBeat — Meta's Muse Ember nearly matches Claude Haiku 5.5 at one-third the estimated per-task cost · AlphaSignal — Meta Releases Muse Ember to Cut AI Agent Costs by 3x

By Shakeel Ahmed — AI Growth Lab

Meta Muse Ember: The $0.10 Model Built to Run Your AI Agents' Grunt Work

Ai Game Changer Pro — Tue, 10 Nov 2026 04:07:09 +0000

Most AI model launches chase the top of the leaderboard. Meta went the other way. On 9 November 2026, Meta Superintelligence Lab released Muse Ember — the lightest and most cost-efficient model in its Muse family, built not to be the smartest model in the room, but to be the one that does the repetitive volume work inside AI agents, at a price of $0.10 per million input tokens.

What Muse Ember is for

Ember is aimed squarely at the unglamorous tasks that make up most of an agent's day:

Tool calls — firing off the many small API actions an agent performs per task.
Coding subtasks — the small, well-defined pieces of a larger coding job.
Routing — deciding which model, tool or workflow should handle each request.
Extraction — pulling structured facts out of unstructured text.
Long-document parsing — chewing through big documents cheaply, page after page.

None of these need a frontier model's full reasoning power. All of them happen thousands of times in a busy agent system — which is exactly where per-token price decides whether the system is viable.

The price: $0.10 per million input tokens

Ember costs $0.10 per million input tokens. Meta positions that as far cheaper than Claude Haiku and OpenAI's Luna, and roughly on a par with the open-source DeepSeek V4.1 Flash and GLM 5.3 Flash models. In other words, a major lab is now pricing a hosted model where the open-source discount tier used to sit — a significant moment for anyone who builds agents on a budget.

Multimodal without a second model

Cheap usually means text-only. Ember is not: it handles text, image and audio in one model — screenshots, charts, scanned documents and voice memos included. For agent builders, that removes a whole layer of plumbing: no separate vision model to call when a subtask involves a screenshot, and no separate transcription step for a voice memo.

Meta's pairing rule: Wonder plans, Ember executes

Meta describes the intended architecture in one line: "Muse Wonder plans, Muse Ember handles the volume." A big, expensive model makes the plan and the rules once; Ember then applies them across thousands of cases. It is the manager-and-team pattern applied to models — you do not pay the senior architect's rate for every brick that gets laid.

When to pick Ember — and when to pick a frontier model

Pick Ember when the subtask is well-defined and repeated: extracting fields from invoices, routing tickets, parsing a backlog of scanned documents, summarising voice memos, or executing the tenth step of a plan a bigger model already made. Volume and predictability are Ember's home ground.
Pick a frontier model when the task is the plan itself: open-ended reasoning, ambiguous problems, high-stakes one-off decisions, or work where a single mistake costs more than a million tokens. Frontier models earn their price on judgement, not throughput.
Mix them, as Meta suggests: let the frontier model decide what should happen and write the rules, then hand the execution loop to Ember. Most real agent systems will end up as this hybrid — the only question is where each builder draws the line.

Bottom line: Muse Ember is not trying to win the intelligence race. It is trying to win the cost-per-task race — and at $0.10 per million input tokens with vision and audio included, it makes a strong case that the future of agents is a cheap, tireless execution layer under a smart planner.

Where to get Muse Ember — and how to check it

Ember is available now in two places: inside Muse Code, and through the Meta Model API under the model ID muse-ember-1.0, with Responses API support. To check it yourself, open Meta's official developer page, dev.meta.ai — Muse Ember is listed there with a Playground where you can try it, alongside the Model API and Muse Code getting-started guides. Full pricing: $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, with a 1 million-token context window.

Source: TechCrunch, "Meta releases Muse Ember, a new lightweight model for agent execution", 9 November 2026 — techcrunch.com. Pricing as reported at launch.

Originally published on Ai Game Changer Pro.

Meta's Muse Ember: Haiku-Level Intelligence at a Third of the Cost — What It Means for Agent Builders

Tiến Thành bình — Tue, 10 Nov 2026 02:19:56 +0000

On November 9, 2026, Meta introduced Muse Ember, the smallest and most efficient model in the Muse family. It is not pitched as a frontier model. It is pitched as the model that does the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.

That framing matters. Most of an agent's token spend is not one brilliant reasoning leap. It is hundreds of small, structured steps. Ember is built for exactly that layer — and Meta's headline claim is that it delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting.

The three-tier lineup

Ember completes a division of labor Meta has been building toward:

Muse Wonder — for the hardest problems.
Muse Spark — for coding and agentic workflows.
Muse Ember — for repeatable coding and agentic tasks at scale.

The line Meta itself uses: "Muse Wonder plans, Muse Ember handles the volume." Ember is also the first model trained using Muse Wonder Full, Meta's largest model — the teacher-student pipeline is now part of the product story, not just a training detail.

Price and specs

	Muse Ember
Input	$0.10 per million tokens
Cached input	$0.004 per million tokens
Output	$0.50 per million tokens
Context window	1M tokens
Max output	512K tokens
Model ID	`muse-ember-1.0`
API	Responses API, Meta Model API, Muse Code

For comparison, Meta lists Muse Spark 1.3 at $1.25 per million input tokens and $4.25 per million output tokens. On Artificial Analysis's cost-per-task measure, Ember comes in at about $0.14 per task versus $2.39 for Spark 1.3 — roughly a 17x gap for a one-point difference on the Intelligence Index (36 vs 37).

Ember is natively multimodal: text, images, audio, and video in, so agents can read screenshots, charts, and scanned documents, or work from voice memos and meeting recordings without a separate model.

The benchmarks: where it wins

All figures below are at maximum reasoning effort, from Meta's official announcement, using Artificial Analysis's November 9, 2026 leaderboard and Meta's own internal runs:

Benchmark	Muse Ember	Claude Haiku 5.5	GPT-6 Luna
Artificial Analysis Intelligence Index	36	36	28
AutomationBench-AA (business workflows)	37.8	13.9	20.2
Agents' Last Exam (end-to-end agent tasks)	61.6	57.0	57.1
Terminal-Bench 4.0 (agentic terminal coding)	40.9	39.2	16.4
Coding Agent Index-AA (real agent harnesses)	46	37	31
OSWorld 2.1 (computer use, partial credit)	69.5	68.9	50.5
PerceptionBench with tools (visual perception)	80.6	78.6	62.8
GDP.pdf-AA (professional long documents)	22.8	20.8	22.8

The headline is the Intelligence Index tie: Ember matches Haiku 5.5's best score at about a third of the cost per task, and beats every Luna setting. Meta also claims Ember sits on the efficiency frontier of the full Artificial Analysis leaderboard — no model scores higher for less per task.

The agent numbers are the more interesting part. AutomationBench-AA, which runs end-to-end business workflows, is not close: 37.8 against 13.9 for Haiku. That is the benchmark that most resembles what Ember is actually sold for — pulling records, calling tools in sequence, updating systems across apps.

Where it trails

The same table shows the limits, and Meta published them rather than hiding them. On FrontierCode 2.0-AA — real engineering tasks judged on mergeability — Ember scores 27.2, behind Haiku 5.5's 34.9. On long-document work (GDP.pdf-AA) it ties Luna at 22.8 and only narrowly beats Haiku.

The pattern is consistent: Ember is strong at executing structured, repeatable agent steps and weaker at the hardest open-ended engineering. That is not a flaw in the pitch; it is the pitch. You do not route your hardest merge to the cheapest model. You route the hundred routine steps around it.

Safety and availability

Before release, Meta evaluated Ember under its Meta Superintelligence Scaling Framework, covering chemical and biological, cybersecurity, and loss-of-control risks. Across the categories assessed, the model remained below the framework's high-risk thresholds, and it received the same core safety training and testing as the larger models. Meta also tested how it behaves when acting on a user's behalf — whether it respects user intent, handles sensitive information appropriately, resists instructions hidden in web pages, documents, or tool outputs, and seeks human input before high-impact actions.

Ember is available today on Muse Code and the Meta Model API as muse-ember-1.0, and on Google Cloud and Oracle Cloud. Meta says more efficient models are already in training.

What it means for agent builders

Three practical takeaways:

1. Route by step type, not by model loyalty. The economical architecture is now explicit in Meta's own lineup: Wonder plans, Spark handles the hard coding workflows, Ember runs the volume. If your agent spends most of its tokens on file reads, test runs, targeted fixes, and extraction, those steps belong on the cheapest model that clears the quality bar — and Ember's benchmark profile says it clears it for exactly those steps.

2. The cost math changes what you can afford to retry. At $0.14 per Artificial Analysis task versus $2.39 for Spark, verification loops, self-correction passes, and multi-sample checks stop being luxuries. A small model that is cheap enough to run three times can beat a big model run once — provided the task is the kind Ember is good at.

3. Check the harness before you trust the number. Meta is transparent that Terminal-Bench 4.0 ran each model in its best published harness (mini-swe-agent for Ember, Claude Code for Haiku, Codex for Luna), and that some results are Meta's own internal evaluations. The direction of the results is consistent across independent and internal benchmarks, but the exact margins depend on setup. Pilot it on your own workload before migrating volume.

The bigger signal is structural: the frontier labs are now competing on the cost of the repeated step, not just the hardest one. Ember is Meta's bet that the agent economy will be won at the bottom of the stack — and at $0.10 per million input tokens, it is a bet other builders can now test for the price of a coffee.

Source: Meta AI Research, "Introducing Our Smallest, Most Efficient Model: Muse Ember," November 9, 2026 — https://research.meta.ai/blog/muse-ember