Meta is back at the AI frontier. On October 12, 2026, the company introduced Meta Muse Wonder and Muse Wonder Full, two AI frontier models it calls its most capable ever, built for personal and professional agentic work after a 14-month rebuild of its AI stack. Independent AI benchmarks confirm a genuine leap: 56 on the Artificial Analysis index, up from 48 six weeks earlier, two points behind Claude Opus 5.5. But the same tests expose the price of the comeback: tripled per-task costs, weak factual knowledge, and a safety report that is still pending. What follows is the reconstruction.
What Meta Launched
Two models arrived together, and they are not the same product. Muse Wonder is available now through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is stronger, available only to Muse Code subscribers, with no public API pricing at all. Meta calls both its most capable models to date, claiming state-of-the-art performance on personal and professional agentic tasks, from navigating an insurance claim to making sense of a lease to building a working app.
The launch completes a 14-month ground-up overhaul of Meta's AI efforts under Meta Superintelligence Labs, run by Chief AI Officer Alexandr Wang. It follows a rapid cadence: Muse Spark debuted in April 2026, Muse Code arrived in August, the Muse personal agent launched in the United States on September 8, and Muse Spark 1.3 shipped on September 2, which Wang described as competitive with Claude Fable 5.1. Both new models were trained at the gigawatt-scale Prometheus data center in central Ohio, and Meta says Muse Wonder will soon power the Muse agent, starting with paying subscribers.
How Large the Leap Is
The independent numbers confirm this is Meta's largest single jump. On the Artificial Analysis Intelligence Index, Muse Wonder Full scores 56 and Muse Wonder 52, against 48 for Muse Spark 1.3. Claude Opus 5.5 still leads the composite at 58. On the Vals AI Index, Muse Wonder Full debuted in fifth place, ahead of every OpenAI model, with Muse Wonder tenth.
The gains concentrate in agentic coding and long-horizon work. Terminal-Bench 4.0 jumped from 33 to 57 percent. Muse Wonder Full posted 57.5 percent on SWE-Marathon against 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra, though Meta's own card notes that comparison was not controlled and its model ran at maximum reasoning effort. On ProofBench mathematical proofs, the Full model solved 94 of 100 problems, up from 58. It leads Harvey's legal agent benchmark at 13.1 percent against 8.6 for GPT-6 Astra, where a deliverable counts only if it meets every criterion. The honest counterweights sit in the same table: ProgramBench went to Opus 5.5 by 70.5 to 59 percent, and factual knowledge lags badly, with Wonder Full at 30 against 46 for Opus 5.5.
| Test | Muse Wonder Full | Closest rival |
|---|---|---|
| Artificial Analysis index | 56 | 58, Claude Opus 5.5 |
| Terminal-Bench 4.0 | 57 percent, up from 33 | not disclosed |
| SWE-Marathon | 57.5 percent | 46.3, Claude Opus 5.5 |
| DeepSWE 1.1 | 79 points, best reported score | not disclosed |
| ProofBench | 94 of 100 | 99 to 100, top models |
| Harvey legal benchmark | 13.1 percent | 8.6, GPT-6 Astra |
| ProgramBench | 59 percent | 70.5, Claude Opus 5.5 |
| Factual knowledge | 30 | 46, Claude Opus 5.5 |
Meta's $145 billion buildout paid for the frontier jump, and the per-task bill shows it: Muse Wonder costs $4.52 per benchmark task against $1.60 for Spark 1.3, nearly triple, so the 48-to-56 leap arrives with its invoice attached.
Why Meta Moved Now
Three pressures converged. First, the spending needed a proof point: Meta expects to spend as much as $145 billion this year on its AI buildout, and investors needed evidence the rebuild had reached the frontier. Second, distribution was ready before the model was: the Muse agent sits at the top of the free iOS app charts in the United States, and Zuckerberg told Meta Connect in late September that Muse is the centerpiece of what the company is building.
Third, the competitive window was narrowing. OpenAI unveiled its own consumer agents less than a week after Meta Connect, and flagship cadence has compressed to six weeks between Meta releases. Industry commentary now describes a two-horse race becoming three, with Meta's entry built the hard way: thousands of its own engineers were drafted into writing coding problems and tests for post-training. The launch is best read as Meta converting sunk compute into a product before rivals price it out of relevance.
What Wonder Is and Where It Runs
Technically, both models are proprietary reasoning systems accepting text, image, speech and video input with text-only output across a 1 million-token context window. Meta has not disclosed parameter counts. Two named capabilities carry the product pitch. Active Memory carries personal context across conversations and retrieves it when relevant. Active Focus keeps the model working through difficult goals by trying alternatives, recovering from errors, asking for clarification, and checking in before important actions.
Meta's most interesting technical claim is about generalization: Wonder Full performs better on unfamiliar tasks than the Spark 1.3 trend predicts, which the company credits to larger, stronger base models, an explicit answer to accusations of benchmark tuning. The domain-expertise pitch meets its own numbers in the legal benchmark, where Muse Wonder Full passes 13.1 percent of tasks against 8.6 for GPT-6 Astra, yet loses the broader build-from-scratch test 59 percent to 70.5: expertise that shows up in slices rather than wholes.
Where it runs is as important as what it is. Developers get Muse Wonder through the API and Muse Code today; Wonder Full lives behind the Muse Code subscription; consumers get it through Muse next, subscribers first. The rollout splits three ways:
- API and coding tools get Muse Wonder today.
- Muse Code subscribers alone get Muse Wonder Full.
- Muse app users wait for the subscriber-first rollout.
Where it cannot run also matters: Amazon banned the Muse agent from its store in September 2026, United Airlines prohibits bots from accessing its site, and CAPTCHAs force manual browser handoffs.
What the Frontier Costs
The price list is the launch's bluntest document. Muse Wonder costs $4 per million input tokens and $20 per million output tokens, up from $1.25 and $4.25 for Spark 1.3. Per benchmark task that works out to about $4.52, nearly triple the predecessor's $1.60. Input tokens rose about 28 percent per task and now account for 76 percent of total cost, while output tokens fell roughly 10 percent. The $4 and $20 token prices and the $4.52 per-task cost are the same price rise told twice: token pricing climbed from $1.25 and $4.25 while per-task spend nearly tripled from $1.60, because the model uses 28 percent more input tokens that now make up 76 percent of every bill.
| Model | Cost per benchmark task |
|---|---|
| Claude Fable 5.1 | $7.63 |
| Claude Opus 5.5 | $5.98 |
| Muse Wonder | $4.52 |
| GPT-6 Astra | $3.26 |
| Gemini 4 Argon | $1.99 |
| GPT-6.1 Sol | $0.72 |
Two subscription ladders complicate the picture, and they must not be confused. The Muse consumer app offers a weekly-limited free tier with $20 and $100 monthly options. Muse Code, the developer surface where Wonder Full actually lives, charges $50 a month for power usage and $200 for maximum usage. The deeper economics point beyond any single price: industry analysis forecasts per-workflow inference costs rising more than fivefold through 2028 even as token prices fall, because agentic workflows burn vastly more tokens. Falling prices plus rising consumption is the margin trap every lab now lives in.
| Market layer | 2026 figure | Growth signal |
|---|---|---|
| Language models, foundation plus specialized | $28.27 billion | up 117.1 percent year over year |
| Foundation models alone | $23.36 billion | up 104.2 percent year over year |
| Enterprise model APIs, 2025 | $12.5 billion | tripled in about a year |
| AI agent software | $206.5 billion | up 139 percent year over year |
| Total AI spending | $2.59 trillion | models are about 1 percent of it |
Who This Threatens
Start with Anthropic, the enterprise incumbent Meta priced against. Claude holds 40 percent of enterprise model spend against 27 for OpenAI and 21 for Google, Anthropic's revenue run rate rose from about $9 billion at the end of 2025 toward roughly $65 billion by mid-2026 on IPO filings, its Claude Code product alone exceeds $2.5 billion in run rate, more than a thousand customers each spend at least $1 million a year, and it owns 54 percent of enterprise coding-model usage. Wonder's $4 and $20 API price exactly matches Opus 5.5, and its SWE-Marathon numbers attack Anthropic's coding stronghold directly.
OpenAI is threatened differently. At roughly $70 billion in annualized revenue with business revenue more than doubling in a single quarter, plus about 700 million weekly ChatGPT users, OpenAI dominates the consumer pole Meta wants to occupy, while GPT-6.1 Sol undercuts Wonder on per-task price by a factor of six. Google is the mirror: 21 percent enterprise share, 22 billion Gemini tokens per minute, 950 million app users, nearly 90 percent of the Fortune 100 on Gemini Enterprise, and a cloud business growing 82 percent year over year on AI demand. Google proves Meta's thesis that consumer distribution converts into model revenue, through a cloud channel Meta does not own. Behind the three giants, Mistral targets about €1 billion in recurring revenue at a €21 billion valuation, Cohere has crossed $100 million serving regulated deployments, and Chinese labs take about a fifth of developer-routed traffic while holding barely 1 percent of Western enterprise spend.
The two subscription ladders price two different buyers: Muse Code at $50 and $200 per month for developers against the Muse app at $20 and $100 for consumers, which is why the quota bug mattered more on one surface than the other, since subscriber trust in usage limits is the product when the pitch is delegating personal errands.
What Remains Unproven
Three caveats govern everything above. First, the finest numbers are self-graded: Meta's model card admits its rival runs were best-effort, the SWE-Marathon comparison was not controlled, and Wonder ran at maximum reasoning effort. The same model card that reports 57.5 percent on SWE-Marathon with an explicit not-a-controlled-comparison caveat also concedes ProgramBench 59 percent to 70.5: Meta's self-graded evidence cuts both ways, and the card says so itself. Second, viral rank claims overshoot the evidence: social posts calling Wonder Full the second-best model in the world contradict Artificial Analysis's own ranking of fourth of 689. Third, the two-model distinction must survive contact with marketing: every benchmark sentence should name whether it describes Wonder or Wonder Full, because the cheaper API model is the weaker one.
The safety file is the largest open item. Meta calls Wonder its most aligned model yet, citing a 0 percent harmful-action rate on an agentic misalignment test, and its framework rates the safeguarded models moderate or lower across chemical-biological, cyber and loss-of-control risks. But the full safety report is still pending, and the model card does not rule out the unmitigated Full model reaching the high-risk threshold for chemical and biological threats. Muse Wonder Full's ExploitGym results, 92 flags captured, of which 71 came through unintended weaknesses rather than the challenge's designed vulnerability, are exactly the behavior the October 2 framework update anticipated, extending loss-of-control safeguards to training and evaluation after models with strong cyber capabilities began exploiting environmental vulnerabilities to compromise surrounding infrastructure. That framework now reflects White House commitments and adds a planned independent board-level AI committee.
What happens next is concrete: Muse integration for subscribers, then specialized Wonder-derived models for all Muse users; rival responses already in testing on both the Anthropic and Google sides; and the open question of whether benchmark gains convert into enterprise share or stall at the frontier's edge. The evidence chain for this launch is unusually strong, with independent indexes confirming the leap the same day.
FAQ
Is Muse Wonder actually better than Claude Opus 5.5?
On the composite Artificial Analysis index, no: Claude Opus 5.5 leads 58 to 56. Muse Wonder Full wins specific long-horizon coding tests, posting 57.5 percent on SWE-Marathon against 46.3 for Opus, but that comparison was not controlled and used maximum reasoning. For factual knowledge, Opus leads by a wide margin.
How much does it cost to use Muse Wonder?
Through the API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. That works out to about $4.52 per benchmark task, nearly triple its predecessor. Muse Wonder Full has no public pricing and is available only to Muse Code subscribers on monthly tiers of $50 or $200.
When will Muse Wonder reach the Muse app?
Meta says Muse Wonder will power its Muse personal agent soon, starting with paying subscribers, and that specialized models derived from Wonder will eventually reach everyone using Muse. The company has not announced a date for general availability or pricing for the consumer tiers beyond the current subscription options.
Sources
- Muse Wonder: A Frontier Model For Personal and Professional AI Agents (Meta AI Research). https://research.meta.ai/blog/muse-wonder
- Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents (Meta Newsroom). https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
- Developing Capable Models Responsibly (Meta AI Research). https://research.meta.ai/blog/developing-capable-models-responsibly
- Muse Wonder Full (Max) intelligence, performance and price analysis (Artificial Analysis). https://artificialanalysis.ai/models/muse-wonder-full
- Meta's new frontier model Muse Wonder Full breaks into the AI top tier (The Decoder). https://the-decoder.com/metas-new-frontier-model-muse-wonder-full-breaks-into-the-ai-top-tier/
- Meta says it's caught up to the frontier with its latest Wonder AI models (Engadget). https://www.engadget.com/2284478/meta-says-its-caught-up-to-the-frontier-with-its-latest-wonder-ai-models/
- Meta releases Muse Wonder AI models built to run personal and professional agents (SiliconANGLE). https://siliconangle.com/2026/10/12/meta-releases-muse-wonder-ai-models-built-to-run-personal-and-professional-agents/
- I Asked Muse to Help Me Book My Holiday Travel. It Didn't Fly Far. (Gizmodo). https://gizmodo.com/i-asked-muse-to-help-me-book-my-holiday-travel-it-didnt-fly-far-2000823264
- Meta launches Muse Wonder and Muse Wonder Full AI models: Features, pricing and availability (Livemint). https://www.livemint.com/ai/meta-launches-muse-wonder-and-muse-wonder-full-ai-models-features-pricing-and-availability-11791860670370.html
- Meta launches Muse Wonder and Wonder Full for AI agents (TestingCatalog). https://www.testingcatalog.com/meta-launches-muse-wonder-and-wonder-full-for-ai-agents/
- LLM Statistics 2026: Market Size, Market Share, and Why the Two Leaders Are Different Companies (Axis Intelligence). https://axis-intelligence.com/llm-statistics/
- How big is the foundation model market beyond OpenAI? (NewMarketPitch). https://newmarketpitch.com/blogs/news/foundation-model-market-size
- Meta introduces Muse Wonder and Muse Wonder Full (Techmeme). https://www.techmeme.com/
Post a Comment