What it is
Muse Wonder is a closed-weights model from Meta Superintelligence Labs, released on October 12, 2026, alongside a stronger variant, Muse Wonder Full. Meta calls them its most capable models and says they sit on larger base models than the Muse Spark line. Muse Wonder (model ID muse-wonder-1.0) is a preview release on the Meta Model API; Wonder Full is not on the API at all and is reserved for Muse Code's $50 and $200 monthly plans.
The model reads text, images, video, audio and PDF and writes text, with a context of 1,048,576 tokens and reasoning effort from minimal to max. Meta pitches it at multi-file code changes, agent runs lasting hours and coordinating sub-agents, and adds two behaviours it calls Active Memory and Active Focus. It shares endpoints and request formats with Muse Spark, so switching is a change of model ID.
API pricing is $4 per million input tokens, $0.20 cached and $20 output, with no long-context premium and no cheaper Contributor mode. It is also listed on OpenRouter at the same rates, and is not offered in South Korea.
Our review
Muse Wonder puts Meta in the frontier tier at exactly Claude Opus 5.5's list price, $4 input and $20 output per million tokens, without matching Opus across the board. On the model card's own table the API model beats Opus on five rows and loses nine. It loses on ProgramBench, 50.0 to 70.5, on Terminal-Bench 4.0, 52.0 to 59.6, and on OSWorld computer use, 72.6 to 81.8. Artificial Analysis puts it at 52 on its Intelligence Index against Opus's 58.
The wins are narrow in subject but real. On DeepSWE 1.1, a long-horizon coding test scored by Datacurve, Wonder reaches 78.8 against Opus's 74.2 and GPT-6 Astra's 74.1. On SWE-Marathon it leads Opus 53.1 to 46.3, though Meta's footnote says that row is not a controlled comparison because rival scores came from Codex and Claude Code. Harvey's legal agent test is lopsided, 10.0 against 4.2, and the internal MemoryBench gives Wonder 85.6 against 81.5, but that benchmark is Meta's own and graded by a model.
The better model is the one most buyers cannot call. Wonder Full scores 57.5 on SWE-Marathon, 59.0 on ProgramBench and 56 on the Artificial Analysis index, yet it exists only inside Muse Code on the $50 and $200 plans. Setting reasoning effort to max on the API does not route to it. A team that wants Full in its own pipeline has no way to buy it.
The terms are tidier than Muse Spark's. There is no Contributor mode for Wonder on the API, so pay-as-you-go traffic is never used for training, while Muse Code subscriptions still default to Contributor for every model, Wonder included. Rate limits match Spark at 3,000 requests and 4 million tokens a minute per team. The model is labelled a preview, and Meta's full safety report was still marked as coming soon at launch.
Where it earns its place
A platform team running a Muse Spark agent pipeline
Send the hardest planning and multi-file steps to a stronger model without a second vendor
Wonder shares Spark's endpoints, so routing a step up is a model ID change. Each step costs 3.2 times as much per input token and 4.7 times per output token, so it pays to keep classification and tool calls on Spark.
An engineering lead at a 30-developer firm evaluating long-running coding agents
Hand off refactors that take hours across many repositories
DeepSWE 1.1 and SWE-Marathon are where Wonder leads Opus 5.5 on Meta's table. A task of 500K input tokens, 400K of them cached, and 50K output costs about $1.48 at list price, the same as Opus, so the choice comes down to which benchmark resembles the work.
A legal-operations manager at a mid-sized law firm
Draft first-pass work products from a matter's documents for lawyer review
Wonder scores 10.0 on Artificial Analysis's version of Harvey's Legal Agent Benchmark against Opus 5.5's 4.2. That is still a small share of tasks passing every check, so output needs a lawyer's review, and API traffic stays out of Meta's training data.
Before you commit
- Setup
- Create a team and API key at dev.meta.ai and call muse-wonder-1.0 in Responses, Chat Completions or Messages format; Muse Code exposes it on every plan.
- Integrations
- Tool calling, parallel tool calls, structured output, desktop computer use and web search grounding at $2.50 per 1,000 queries; listed on OpenRouter at the same token prices.
- Data and privacy
- Standard mode only on the API: prompts, files and outputs are not used to train Meta models. Muse Code subscriptions default to Contributor mode, which does train on content, until switched off. Closed weights, no self-hosting.
- Cost at scale
- A million support replies a month at 3,000 input and 500 output tokens each is about $12,000 input plus $10,000 output, $22,000, against about $5,875 on Muse Spark 1.3. Cached prefixes at $0.20 cut the input share.
- Limits
- 1,048,576-token context; 3,000 RPM and 4M TPM per team; preview release; not available in South Korea. Meta's card names visual and 3D spatial reasoning as a known weakness.
What's good
- Leads Opus 5.5 on DeepSWE 1.1 (78.8 to 74.2) and SWE-Marathon (53.1 to 46.3)
- 1M-token context with no long-context premium; cached input at $0.20
- No Contributor mode on the API, so API traffic is not used for training
- Same endpoints as Muse Spark; a model ID change moves work between them
What's not
- Loses 9 of 14 rows on Meta's own table to Opus 5.5, at the same price
- Wonder Full, the stronger variant, is not sold on the API
- Preview status; full safety report not out at launch
- Muse Code subscriptions default to Contributor mode, which trains on your content
Best for
Teams already on Meta's API that need a stronger model for long, multi-file coding or legal document work than Muse Spark provides.
Alternatives
Muse Wonder: common questions
- Is Muse Wonder worth it?
- Wonder costs what Claude Opus 5.5 costs and wins only five of the fourteen rows Meta chose to publish against it. We rate it 4.1 out of 5.
- How much does Muse Wonder cost?
- API: $4 input, $0.20 cached input, $20 output per 1M tokens; no Contributor mode. Wonder Full only in Muse Code Power ($50/month) and Maximum ($200/month).
- What are the downsides of Muse Wonder?
- Loses 9 of 14 rows on Meta's own table to Opus 5.5, at the same price; wonder Full, the stronger variant, is not sold on the API; preview status; full safety report not out at launch; muse Code subscriptions default to Contributor mode, which trains on your content.
- Who is Muse Wonder best for?
- Teams already on Meta's API that need a stronger model for long, multi-file coding or legal document work than Muse Spark provides.
- How much does Muse Wonder cost for a team or in production?
- A million support replies a month at 3,000 input and 500 output tokens each is about $12,000 input plus $10,000 output, $22,000, against about $5,875 on Muse Spark 1.3. Cached prefixes at $0.20 cut the input share.
- Is Muse Wonder safe for company data?
- Standard mode only on the API: prompts, files and outputs are not used to train Meta models. Muse Code subscriptions default to Contributor mode, which does train on content, until switched off. Closed weights, no self-hosting.
- What are the best Muse Wonder alternatives?
- The closest alternatives we have reviewed are Claude Opus 5.5, GPT-6 Astra, Claude Fable 5.1 and Muse Spark 1.3.
Answers checked Oct 23, 2026.
Make Muse Wonder? Show this rating on your site
