Meta Muse Wonder, Explained: What It Scores, What It Costs, and Why the Best Version Is Not on the API

What Meta released

On October 12, 2026, Meta Superintelligence Labs released two models: Muse Wonder and the more capable Muse Wonder Full. TechCrunch called them Meta's first attempt at a frontier model since it rebuilt its AI effort in August 2025, and noted that both were trained at Meta's Prometheus data center in Ohio. Meta's own line is that "they mark the next step on our scaling ladder." They are a larger class than the Muse Spark models Meta has shipped since April, which stay on sale as the cheap option.

Meta's developer post pitches both at the same jobs: "frontier coding, long-running agents and advanced computer use." Both have a one million token context window and take screenshots, charts, documents, images and video as input. Meta says audio support still has to be confirmed on each product.

The catch: two models, two ways in

The model you can build on is not the one that scores highest.

• Muse Wonder is on the Meta Model API as muse-wonder-1.0. It costs $4 per million input tokens, $0.20 for cached input and $20 for output, and VentureBeat reports $2.50 per 1,000 searches when it uses web search. The API speaks the OpenAI format, so existing code needs a new key from dev.meta.ai, a new base URL and the new model name.
• Muse Wonder Full has no per-token price at all. It comes only with Muse Code, Meta's coding agent, on the Power Usage plan ($50 a month) and the new Maximum Usage plan ($200 a month, with 80 times the allowance of the entry plan).

Meta says Wonder brings "comparable performance to Muse Wonder Full." The independent numbers put the gap at four points on the index, which is the difference between the middle of the frontier pack and its top.

How it scores on the independent index

Artificial Analysis runs every model through the same set of evaluations and folds them into one Intelligence Index. These are its launch-day scores as VentureBeat reported them, with each model's API price:

Two readings matter. Wonder Full is genuinely at the frontier: level with Sonnet 5.5 and ahead of GPT-6 Astra, behind only Opus 5.5. Wonder, the one on the API, lands just behind Astra and Gemini 4 Argon, at Opus 5.5's per-token price. Per token is not the whole bill: Artificial Analysis put Wonder's cost for an average index task at $4.52, more than GPT-6 Astra's $3.26 and less than Opus 5.5's $5.98, because the models use different numbers of tokens to finish the same job.

Where Meta's own numbers are strongest: long coding jobs

Meta leads its launch with coding. On the two long-horizon software benchmarks it published, both Wonder models come out ahead of Opus 5.5 and GPT-6 Astra:

Read the table whole. Wonder wins the long multi-step coding runs, and loses Terminal-Bench, the agent-in-a-shell test, and the medical benchmark to Opus 5.5 and Astra. VentureBeat also cautions that Meta's table mixes third-party runs with Meta's own and with vendor-supplied scores, and that some coding comparisons used different harnesses. Treat the coding lead as a reason to try it on your own repository, not as settled.

Your data, and which mode you are in

This is the part to check before you sign up. VentureBeat reports that Muse Code subscriptions default to Contributor mode, which lets Meta use what you send to improve its models. The API runs only in Standard mode. Qualifying API customers can get zero data retention, but it switches off server-side conversation state, background requests, file uploads and built-in web search. Neither Wonder model has downloadable weights, so there is no self-hosting.

Which one to use

Both models also power Meta's personal agent, covered in Meta Muse, explained. For the two Claude models it is measured against, see Claude Sonnet 5.5 vs Opus 5.5, and for what every model costs, the LLM pricing comparison. The listings are at Muse Wonder and Muse Spark 1.3, and the launch is in the October 12 Daily Ten.