What Meta Muse Wonder launched on October 12
Meta Muse Wonder arrived on Monday, October 12, 2026, as the company’s first frontier-class models, and developers noticed the price before anything else. The API-accessible model, Muse Wonder, is available through the Meta Model API and Muse Code. Its stronger sibling, Muse Wonder Full, is limited to paid Muse Code subscribers. Meta calls both its most capable models to date and says they are built for agentic work, such as handling an insurance claim, reviewing a lease or building an app.
Both models are set to power Meta’s Muse personal agent app, starting with subscribers on the $20 and $100 monthly plans, according to Futurum Group. Meta says the models include Active Memory, which carries information across conversations, and Active Focus, which keeps a task moving through errors and asks for clarification or permission before acting, as Yahoo Finance reported. Meta also says it trained the models at its Prometheus data center in Ohio, which came online this year, per Yahoo Tech.
The launch matters because it is Meta’s first frontier release since Scale AI founder Alexandr Wang restructured the company’s AI efforts in August of last year. Since then Meta has shipped the lightweight Muse Spark model and the Muse agent, which launched in early September, but it had avoided the state-of-the-art releases that OpenAI and Anthropic are known for. Muse Wonder is Meta’s first real attempt to compete at the top of the frontier market rather than near it.
Meta Muse Wonder price and API access
Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, with a 1-million-token context window, according to MLQ.ai and Technology.org. At list price, a workload of one million input tokens and one million output tokens costs $24. Cached input at $0.20 is a 95 percent discount on the standard input rate, so applications that resend the same long prompts can cut costs sharply.
Pricing is where Meta’s message gets tangled. Alexandr Wang told CNBC the model would be priced “meaningfully below” the leading OpenAI and Anthropic models, as Felloai reported. Yet Muse Wonder’s token prices match Anthropic’s Claude Opus 5.5, and the launch coverage does not explain the gap between that claim and the list price. Analysts at Futurum describe the frontier market as converging, with similar capabilities and the $4 and $20 price point becoming a standard. If that holds, price alone will not decide which model developers pick.
Which Muse Code plans unlock Wonder Full
Muse Wonder Full is sold only through Muse Code’s $50-a-month plan, called Power Usage, and its $200-a-month plan, called Maximum Usage. Meta lists no API price for it, so developers cannot call the stronger model the way they can call Muse Wonder. Anyone who wants Meta’s top model needs a subscription first.
That split creates clear winners and losers. Developers can reach Muse Wonder without a subscription, while consumers must commit at least $50 a month to get the strongest model Meta has shipped. Wells Fargo said Meta has reached the “AI frontier,” citing the Artificial Analysis index to place Wonder Full second only to Anthropic’s Claude Opus 5.5, as Quartz reported. That ranking makes the paywall easier to justify for heavy users who can spread the cost across paid work.
How Meta’s benchmarks compare with Claude Opus 5.5 and GPT-6 Astra
Meta’s comparisons pit Muse Wonder against Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra. Other comparisons in coverage also include GPT-6.1 Sol and Claude Fable 5.1. The table below combines Meta’s own tests with the independent Artificial Analysis Intelligence Index.
| Benchmark | Muse Wonder | Muse Wonder Full | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Meta MemoryBench (Meta’s test) | 85.6 | 85.5 | 81.5 | 81.4 |
| Meta DeepSWE 1.1 (Meta’s test) | Not cited | 79.0 | 74.2 | 74.1 |
| Artificial Analysis Intelligence Index | 52 | 56 | 58 | Not cited |
On Meta’s MemoryBench, the standard Muse Wonder edges out Wonder Full by a tenth of a point, a reminder that the stronger sibling does not win every row. On DeepSWE 1.1, Wonder Full leads Opus 5.5 by 4.8 points. Those margins are small, and both come from Meta’s own testing.
Zuckerberg said on the Cheeky Pint podcast, “now everything that we do going forward will just kind of be pushing the frontier.” As CNBC reported through Quartz, he also said the time between successive model releases should shorten. Wang called the models built for personal and professional agentic tasks and said, “These are true frontier models.” Those statements describe ambitions. The benchmark tables are the evidence, and that evidence is narrower than the rhetoric.
Where independent scores fall short
The independent picture is less flattering. On the Artificial Analysis Intelligence Index, Muse Wonder scores 52, Wonder Full 56 and Claude Opus 5.5 58, according to MLQ.ai and Felloai. Wonder Full trails Opus 5.5 by two points, and standard Wonder trails by six. Neither Meta model takes the top spot on that index, which conflicts with the company’s framing.
Meta compiled most of the benchmark results itself, and independent checks were not published as of launch coverage. Accounts also disagree on how many of Meta’s own benchmark rows Muse Wonder wins. One says it leads on 7 of 14 rows, and another says it beats Opus 5.5 on 5 of 14. The 7-of-14 figure traces to a secondary blog that we could not verify, so treat it as unconfirmed.
Safety documentation and openness are also incomplete. Meta says a fuller safety report is coming, so the material available at launch is partial. The models have no open weights, and the Muse Spark weights Meta promised in September have not shipped, according to Felloai. Outside researchers cannot download and run the models locally, which limits independent checks to what the API and subscription products reveal.
What developers and subscribers should do now
If you build agents or depend on Meta’s models, run your own workload before switching:
- Test Muse Wonder against Claude Opus 5.5 and GPT-6 Astra on your real tasks, since independent index scores sit below Meta’s tables.
- Model cost with your own mix of input, cached input and output tokens, because an output-heavy agent costs far more per call.
- Test the 1-million-token context window with your largest documents instead of assuming long inputs behave well.
- Check how often your work truly needs Wonder Full before choosing between the $50 Power Usage plan and the $200 Maximum Usage plan.
- Expect Muse app features to reach $20 and $100 subscribers first, and check Meta’s rollout notes before planning around them.
- Wait for Meta’s full safety report and independent testing before relying on the models for sensitive work such as insurance claims or lease reviews.
What investors and rivals should watch next
The stock reacted quickly. TIKR reported a gain of about 3.4% to close at $742.96 on Monday, and a Charles Schwab market update cited a 2.6% gain on Tuesday. Figures differ by the day measured and across sources, so verify before quoting them. TIKR’s analysis asks whether the model can pay for Meta’s $89 billion in capital expenditure. Whether growth in the Muse agent will translate into revenue remains unknown.
Anthropic and OpenAI now face a rival that matches Opus 5.5 on list price while claiming frontier results. Developers will settle that comparison with their own test runs, not with Meta’s tables. If the $4 and $20 price point becomes the norm, competition may shift toward speed, reliability and access terms.
Zuckerberg’s prediction that release gaps will shrink sets the next test. Watch for Meta’s full safety report, fresh runs of the Artificial Analysis index, any change to Muse Wonder’s list price, and whether Wonder Full reaches the API. The Wonder models will eventually power more of the Muse agent, so the coming months will show whether Meta’s subscription bet holds.
