Meta's Muse Wonder Full Scores 56 on a Key AI Index. You Can't Buy It by the Token

Meta has released two new AI models, Muse Wonder and Muse Wonder Full, and says they deliver "state-of-the-art performance on personal and professional agentic tasks." The stronger one scores 56 on the independent Artificial Analysis Intelligence Index. By the figures in Meta's own model card, that puts it above Claude Fable 5.1 and GPT-6 Astra (53 each) but two points behind Claude Opus 5.5 (58).

There is a catch in how you get it. Developers can pay per token for Muse Wonder, at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full, the model behind the headline numbers, is only available through a Muse Code subscription.

Based on Meta's research blog post, newsroom post and model card, Artificial Analysis's published scores, and reporting from CNBC, VentureBeat, TechCrunch and The Verge, checked October 13, 2026. We have not tested either model. Benchmark figures are Meta's or Artificial Analysis's.

What Meta released, and how the two versions differ

Muse Wonder and Muse Wonder Full are the first models in what Meta calls the Muse Wonder series. They follow Muse Spark, which The Verge describes as the first model from Meta's revamped Superintelligence Labs. Meta says both were trained at its "gigawatt-scale Prometheus data center" and build on "larger, stronger base models" than Muse Spark 1.3.

Meta does not publish parameter counts or architecture details for either model. What it does show is a performance gap. Muse Wonder Full beats Muse Wonder on most benchmarks in Meta's own table. Meta also says Wonder Full handles unfamiliar tasks better than its in-house trend would predict.

According to the Muse Wonder model card, both models take text, image, video and audio as input, reply in text, and have a context length of "1M tokens at inference." VentureBeat reports that neither model is being released with downloadable weights. That means you cannot self-host either one.

Who can use it, and what it costs

Meta's newsroom announcement says Muse Wonder is "available today for developers and builders through the Meta Model API and Muse Code," and Muse Wonder Full is "available today for Muse Code subscribers." In summary:

• Muse Wonder via the Meta Model API: $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, according to Artificial Analysis.
• Muse Wonder Full via Muse Code: CNBC reports the coding agent costs $50 or $200 a month. Meta has not announced per-token pricing for Wonder Full.
• Muse, the consumer agent app: Meta says Wonder "will soon power experiences in Muse" and will reach subscribers first. It has given no date. CNBC reports that Muse plans cost $20 or $100 a month.

Here is how the API price works out in practice. On our own calculation, a request with 100,000 input tokens and 10,000 output tokens would cost about $0.60 on Muse Wonder: $0.40 for input and $0.20 for output. Agent tasks can cost far more, because they resend context many times. Artificial Analysis measured about 7.69 million input tokens per index task for Muse Wonder.

Artificial Analysis estimates that one task in its index costs $4.52 on Muse Wonder (max). That compares with $5.98 on Claude Opus 5.5 and $3.26 on GPT-6 Astra, at the settings it tested.

Active Memory and Active Focus, in Meta's words

Meta is promoting two capabilities for agent work. The research post defines Active Memory as "carrying important context about you across conversations and actively retrieving the details relevant to the task at hand."

Active Focus, per the same post, describes how Muse Wonder "continues working toward a challenging goal: trying alternative approaches, working through blockers, and adapting to new information." Meta adds that the model "asks for clarification when needed and checks in before taking important actions."

Speaking to CNBC, Meta AI chief Alexandr Wang described the aim: "We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time." Meta scores these abilities on internal tests called MemoryBench and FocusBench. Outsiders cannot run either one.

What the benchmarks say, and who ran them

Meta's model card compares the two Wonder models with Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 on 14 benchmark rows, all at max reasoning effort. Our count of that table: Muse Wonder Full posts the top score on six rows, Opus 5.5 on six, Muse Wonder on one (MemoryBench) and GPT-6 Astra on one.

Wonder Full's wins are concentrated in long-running and professional agent work. They include DeepSWE (79.0) and SWE-Marathon (57.5, run at "xhigh" effort), plus legal, finance and business-automation tasks. Opus 5.5 leads in Meta's table on Terminal-Bench, ProgramBench, OSWorld, HealthBench Pro, JobBench and the Artificial Analysis index.

The Artificial Analysis listing independently puts Muse Wonder Full (Max) at 56 and Muse Wonder (Max) at 52. Both are up from 48 for Muse Spark 1.3.

The catch: what is still unverified

• Mixed benchmark sources. Meta's methodology says it uses third-party results where they exist. Otherwise it uses "the highest of our internally validated run, self-reported score, or official leaderboard score." Meta also admits that competitor runs "may not be optimally tuned for proprietary models."
• Not like-for-like everywhere. Meta notes that the SWE-Marathon row "is not a controlled common-harness comparison," because rivals were scored with Codex and Claude Code.
• Internal tests. Meta's headline Personal Superintelligence Index, MemoryBench and FocusBench are its own evaluations.
• No price for the top model. You cannot yet buy Wonder Full by the token, so its true cost per task is unknown.
• Safety report pending. Meta calls Wonder its "most aligned model to date," but says a full safety report "will be available soon."
• Data use. VentureBeat reports that Muse Code subscriptions default to a mode that lets Meta train on submitted content. It says you can switch this off, but only for future submissions.

The model card also lists known weaknesses, including "Visual and spatial reasoning, especially 3D tasks," and errors in multi-step reasoning in novel scenarios.

Frequently asked questions

Can I use Muse Wonder Full through the API?

Not yet. Meta says Wonder Full is available to Muse Code subscribers, and Artificial Analysis notes that per-token pricing for it has not been announced. Only Muse Wonder is on the Meta Model API.

Is Muse Wonder better than Claude or GPT?

It depends on the task, and the figures come from Meta or Artificial Analysis. On the Artificial Analysis index, Wonder Full (56) sits above Fable 5.1 and GPT-6 Astra but below Opus 5.5 (58), according to Meta's own table. Test it on your own workload before switching.

Are the models open source?

No. VentureBeat reports that neither Wonder model ships with downloadable weights, so access is only through Meta's API and apps.

For developers, the practical first step is to try Muse Wonder on the API with a capped budget and compare it with your current model on your own tasks. Wonder Full's headline numbers will be hard to judge until Meta prices it by the token and publishes its full safety report.