Muse Wonder: Opus-level prices, and Meta keeps the better model off the API
Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026, and updated the post on October 13. Meta calls them "our most capable models to date". Artificial Analysis calls the release "a material jump in intelligence over previous Meta models", and for the first time a Meta model sits in the same range as the frontier labs on its index. You can use Muse Wonder today through the Meta Model API and Muse Code. Muse Wonder Full, the stronger of the two, is only available to Muse Code subscribers.
For anyone shipping software, the headline is the business model, not the benchmark. Meta priced the API model exactly like Claude Opus 5.5. It kept the better model behind a $50/month coding-CLI plan, and it released no weights for either one. That is a long way from the Llama-era Meta.
What actually shipped
Muse Wonder (muse-wonder-1.0) is available on the Meta Model API and in Muse Code. The model card lists a 1M-token context window. Inputs are text, image, video and audio; output is text only. The terms allow commercial and research use under the Meta Model API Terms of Service and Acceptable Use Policy. Muse Wonder Full is Muse Code only, and the Muse Code plans page names it only from the Power Usage tier ($50/month) upward.
Both come with agent features. "Active Memory" carries context across conversations. "Active Focus" keeps working toward a goal, recovers from errors and checks in before important actions.
There are no weights. Artificial Analysis lists Muse Wonder as proprietary. If you need something local from Meta, look at the separate Muse Glimmer line.
On safety, Meta says a full report "will be available soon". The model card already has safety numbers, including 88.4% on CyberGym for Wonder Full. Meta also writes that it could not rule out high-risk Chem/Bio capability before mitigations.
Numbers that matter
Meta's capability claims come as a chart and adjectives ("state-of-the-art" on Harvey's LAB, FocusBench and SWE-Marathon). The newsroom post and the model card give no head-to-head scores in text. That makes Artificial Analysis the most useful source here.
| Muse Wonder | Muse Wonder Full | Source | |
|---|---|---|---|
| AA Intelligence Index v4.3 (max) | 52 | 56 | Independent (AA) |
| Terminal-Bench 4.0 | 52% | 57% | Independent (AA) |
| AA Coding Agents Index | 63 | 68 | Independent (AA) |
| AA hallucination rate | 27% | 23% | Independent (AA) |
| Input / cached / output, $ per M tokens | $4 / $0.20 / $20 | not on API | Meta pricing page |
| Cost per AA Index task | $4.52 | n/a | Independent (AA) |
For context, AA reports per-task index costs of $5.98 for Claude Opus 5.5, $3.26 for GPT-6 Astra and $1.99 for Gemini 4 Argon (high). Its Coding Agents Index puts Claude Code with Sonnet 5.5 at 68, level with Wonder Full, and Opus 5.5 at 66. AA's model page shows 25.2 s to first token and 145 tokens/s output on Meta's API.
Here is an illustrative monthly cost (my workload, not Meta's): an agent fleet using 2B input tokens a month, 80% of them cache hits, plus 200M output tokens (reasoning included, since reasoning bills as output).
| Line | Calculation | Cost |
|---|---|---|
| Cached input | 1.6B × $0.20/M | $320 |
| Uncached input | 0.4B × $4/M | $1,600 |
| Output | 0.2B × $20/M | $4,000 |
| Total | about $5,920/month |
Meta charges nothing for cache writes. Meta's pricing page also lists web-search grounding at $2.50 per 1,000 queries on top of the token cost.
Meta priced the API model exactly like Claude Opus 5.5, kept the better model behind a $50/month coding-CLI plan, and released no weights for either one.
Technical meaning and trade-offs
The best Meta model is not something you can build products on. Wonder Full exists only inside the Muse Code CLI. The API model scores four index points lower and five points lower on Terminal-Bench 4.0. If your evaluation was run in Muse Code, it probably used Full. Your production integration would get Wonder.
The data terms differ by product. On the API, Meta states "Contributor mode isn't available for Muse Wonder". The page also says Muse Code subscriptions default to Contributor mode, which lets Meta train on your prompts, code and files, and you switch it off in settings. That matters more than the price for anyone pointing a CLI at a client repo.
Then there are latency and limits. A 25 s TTFT at max effort is fine for background agents and poor for interactive UIs. Rate limits are 3,000 RPM and 4M TPM, counted per team rather than per key.
Hallucination is the weak spot. On AA's numbers, Wonder Full hallucinates at 23% and Wonder at 27%, against 15% for Gemini 4 Argon. For legal and finance work, the domains Meta markets most heavily, that gap is the number I'd watch.
My read: Meta built a competitive model and declined to undercut the market leader on price. Buyers will need another reason to pick it, such as memory features, audio input, or wanting a second vendor.
What I'd do
First, evaluate through the API, not Muse Code, so you are testing the model you would actually deploy. Second, run it as a second provider behind your existing router for long-running agent tasks, and compare cost per completed task, not per token. AA's figures ($4.52 vs Opus 5.5's $5.98 per index task) suggest Wonder can come out cheaper on agent work despite the same list price.
Third, turn off Contributor mode on any Muse Code seat that touches proprietary code, and write the setting into your onboarding checklist. Fourth, gate legal and finance use on your own hallucination evals; Meta's "state-of-the-art on Harvey LAB" claim has no published number behind it yet. Fifth, don't plan around weights. If you need self-hosting or data residency, Muse Wonder doesn't meet that requirement. See Muse Spark and Muse Agent for how Meta has split its open and closed lines.
What I couldn't verify
Meta's own benchmark scores appear only in charts. I found no figures for LAB, FocusBench or SWE-Marathon in the text, so every number in the table above comes from Artificial Analysis. Claims attributed to Meta by secondary sites, such as an AA index comparison "from Meta's launch post" and a "Deep Suite 79%" figure, I couldn't trace to a Meta page, so I left them out.
The full safety report isn't published. The model card's ExploitGym figure was revised on October 23 (from 21 to 20 intended flags), so expect more edits. Parameter count and architecture are undisclosed. For Opus 5.5 price parity, I took Opus 5.5's rates from secondary coverage (eesel.ai), not from Anthropic's pricing page in this pass. For Muse Wonder Full in Muse Code, I took the Power-tier gating from the plans page; usage caps per plan are given as relative multiples, not tokens.
I'll revisit when the safety report lands or Wonder Full reaches the API.
Sources
Primary: Meta Newsroom: Introducing Muse Wonder models; Meta Research: Muse Wonder blog; Muse Wonder model card; Meta Model API: Pricing and rate limits; Muse Code plans. Independent: Artificial Analysis: Muse Wonder and Muse Wonder Full; Artificial Analysis: Muse Wonder model page. Secondary: eesel.ai: Meta Muse Wonder pricing (secondary, Opus comparison).