Meta has released Muse Wonder and Muse Wonder Full, two new models from Meta Superintelligence Labs built for agents that act on your behalf at work and in daily life. Independent testing puts both near the top of the frontier, and Muse Code with Wonder Full now matches Claude Code with Sonnet 5.5 on coding agent tests. The catch for builders is simple: the stronger of the two, Muse Wonder Full, is not available through the API at all.
The benchmark figures below come from Meta and from Artificial Analysis, which tested both models independently. 9to6AI has not run either model yet.
Key points
- What shipped: Muse Wonder on the Meta Model API and in Muse Code, and Muse Wonder Full only inside Muse Code.
- Independent scores: Artificial Analysis gives Muse Wonder Full 56 and Muse Wonder 52 on its Intelligence Index, up from 48 for Muse Spark 1.3, which shipped about six weeks earlier.
- API price: $4 per million input tokens, $0.20 cached and $20 output. That is roughly three times Muse Spark 1.3’s input price and almost five times its output price, and it matches Claude Opus 5.5.
- Specs: a 1M-token context window, with text, image, video and audio input and text output.
- Still missing: Meta’s full safety report, which it says is coming soon. There is no word on open weights.
What Meta shipped
Meta’s announcement calls the pair its “most capable models” and frames them as the next step toward what it calls personal superintelligence. Both were trained at its gigawatt-scale Prometheus data center. Meta says the models are tuned for agent work in four areas:
- Domain expertise: working with source documents in law, finance and health. Meta cites Harvey’s Legal Agent Benchmark as one of its tests.
- Active Memory: carrying important context about a user across conversations and pulling it back in when it becomes relevant.
- Active Focus: keeping at a hard task by trying other approaches and working through blockers, while checking in before important actions.
- Frontier coding: long-running software work, measured with Meta’s own SWE-Marathon test.
Meta’s demos show the model planning a running schedule inside a browser, designing and ordering custom merchandise from Instagram photos, and building a small 3D kitchen game. These are vendor demos, not independent tests.
Muse Wonder will also come to the consumer Muse app, starting with subscribers. Meta adds that it will use Wonder to build smaller, specialized models for everyone else.
What the independent numbers say
Artificial Analysis found that Muse Wonder Full beat Muse Spark 1.3 on all ten evaluations in its index. The biggest gains were in agent tasks:
- Terminal-Bench 4.0: 33% to 57% for Wonder Full, which puts it level with Gemini 4 Argon. Wonder reached 52%.
- GDPval-AA v2.1 (real-world knowledge work): 1857 Elo for Wonder Full, second place and within the margin of error of Claude Opus 5.5 at 1866.
- Coding agents: Muse Code with Wonder Full scores 68 on the Coding Agents Index, the same as Claude Code with Sonnet 5.5 and above Opus 5.5 at 66. Muse Wonder scores 63, and the previous Muse Code setup scored 54.
- Hallucination: Wonder Full’s rate fell from 33% to 23%. Among models scoring 45 or more on the index, only Gemini 4 Argon hallucinates less.
There are weak spots too. General knowledge is still behind the leaders: on AA-Omniscience, Wonder Full scores 30, while Claude Opus 5.5 scores 46 and GPT-6 Astra 43. Meta’s memory claims also rest mostly on its own tests. Its MemoryBench measures recall over conversations of 64,000 to 100,000 tokens with AI grading, which says little about remembering a real user over months.
The price jump is the real signal
Earlier Muse models undercut rivals on price. This one doesn’t. Muse Spark 1.3 cost $1.25 in and $4.25 out per million tokens, and Muse Wonder costs $4 in and $20 out. That is double the price of Claude Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol, the same as Claude Opus 5.5, and 60% below GPT-6 Astra and Claude Fable 5.1.
The model also reads a lot. Artificial Analysis measured about 7.7 million input tokens per index task, so a full run costs $4.52 per task. That is cheaper than Opus 5.5 or Sonnet 5.5 at max effort, but more than six times GPT-6.1 Sol at $0.72 and nearly three times Muse Spark 1.3. Prompt caching matters a lot here: cached input is 95% cheaper.
What it means for builders
The model you can call and the model that tops the charts are not the same. Muse Wonder Full is only in Muse Code. Reports on Meta’s availability table say it needs the Power Usage plan at $50 a month or Maximum Usage at $200 a month, with no API key route. If you are building a product, you are choosing Muse Wonder, the one that scores 52.
A few practical calls:
- Agent and knowledge-work pipelines: Wonder is worth a side-by-side test against Opus 5.5 and Sonnet 5.5 on your own tasks, especially long document work, where its low hallucination rate may help.
- Cost-sensitive, high-volume work: GPT-6.1 Sol and Gemini 4 Argon still cost far less per task. Wonder’s heavy input use will show up on your bill unless you cache well.
- Coding: Muse Code with Wonder Full is now a real alternative to Claude Code on paper. Early users on Hacker News say the Muse Code sandbox has been unreliable on Mac and Linux, so try it on a side project before moving a team over.
- Personal-data agents: Meta pitches Wonder for handling messages, calendars and private documents. Wait for the full safety report before you give it that kind of access in production.
What to watch
- Whether Muse Wonder Full reaches the API, and at what price.
- The full safety report, especially how often the model mishandles private information.
- Independent memory tests over long, real usage rather than single long conversations.
- Responses from OpenAI and Anthropic now that a third lab is pricing at the top of the market.
