Meta Superintelligence Labs released two models on October 12, 2026, and developers can call only one of them. Muse Wonder is available through the Meta Model API and in Muse Code, Meta’s coding agent. Muse Wonder Full, the larger model and the stronger one in Meta’s own tests, runs only inside Muse Code on the $50 and $200 monthly plans; Meta’s model documentation lists it as not available on the API. Most results below therefore come in two parts: what Full shows the family can do, and what Wonder delivers in software you can build yourself.
In Meta’s benchmark results, Full is strongest on long software projects, legal agent work, and business workflows, and trails Claude Opus 5.5 on from-scratch programming, terminal coding, health questions, and computer use. On the independent Artificial Analysis Intelligence Index, Opus 5.5 scores 58, Full 56, and Wonder 52. Wonder costs $4 per million input tokens and $20 per million output tokens, the same token rates as Opus 5.5. We have not tested either model.
Muse Wonder at a glance
- Released: October 12, 2026, by Meta Superintelligence Labs
- Muse Wonder: model ID muse-wonder-1.0, in preview; Meta Model API and every Muse Code plan
- Muse Wonder Full: Muse Code Power Usage ($50 a month) and Maximum Usage ($200 a month) only
- API price: $4 input, $0.20 cached input, $20 output per million tokens; no minimum spend, no long-context premium
- Context and inputs: 1,048,576 tokens; text, image, video, audio, and PDF in, text out
- Independent score: Full 56, Wonder 52 on the Artificial Analysis Intelligence Index (October 12, 2026)
- Weights: proprietary; no open weights announced
What is new in Muse Wonder?
1. Two models, and whether Meta trains on your work
Muse Spark, released in April 2026, was the lab’s first model family. Zuckerberg calls Muse Wonder a much larger class of models that brings Meta to the frontier: Muse Wonder for fast frontier performance, Muse Wonder Full as the full version.
Moving from Muse Spark to Muse Wonder on the API is a model ID change, since both take the same requests on the same endpoints. Full is reachable only through the Muse Code command-line agent, and setting reasoning effort to max on Wonder does not select it.
The data terms differ by route. On the API, Wonder runs only in Standard mode, where Meta says your content is not used to train its models. Muse Code subscriptions, including the plans that unlock Full, default to Contributor mode: more usage in exchange for letting Meta train future models on your prompts, code, files, and the responses you get back. It can be turned off in settings at any time, which is worth doing before sending work code.
2. Price: Opus 5.5 token rates, several times Muse Spark
| Per million tokens | Muse Wonder | Muse Spark 1.3 (Standard) | Claude Opus 5.5 |
|---|---|---|---|
| Input | $4.00 | $1.25 | $4.00 |
| Cached input | $0.20 | $0.15 | $0.20 |
| Output | $20.00 | $4.25 | $20.00 |
Source: Meta Model API pricing, October 12, 2026; Anthropic’s Opus 5.5 rates at its September 22 launch.
Equal token rates do not mean equal bills. In Artificial Analysis’s runs of its Intelligence Index, Wonder cost $4.52 per task, against $5.98 for Opus 5.5, which scored 6 points higher, and $1.60 for Muse Spark 1.3, which scored 4 points lower. That is 2.8 times Spark’s cost on one test suite, not a multiplier for your application: the figure is a weighted average over that suite, and a real bill depends on input volume, caching, and how many steps an agent takes.
3. Active Memory and Active Focus
Active Memory is Meta’s name for carrying important context about you across conversations and retrieving what matters for the current task. Active Focus means continuing toward a hard goal: trying other approaches when blocked, asking for clarification, and checking in before important actions.
Both are measured on Meta’s internal tests. On MemoryBench, which checks recall across conversations of 64,000 to 100,000 tokens with distractors and later corrections, Wonder scores 85.6 and Full 85.5, against 81.5 for Opus 5.5 and 81.4 for GPT-6 Astra. On FocusBench, personal tasks in housing, travel, IT, and finance with real-world blockers, Full scores 54.2% against 47.7% for GPT-6 Astra and 45.8% for Opus 5.5; Wonder scores 44.9%.
4. Where Full leads and where it trails
Meta compares the two Muse models with Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1 on 14 benchmarks, all at max reasoning effort. The rows below show the pattern; the full table is in Meta’s announcement and model card.
| Benchmark (what it tests) | Wonder Full (max) | Wonder (max) | Claude Opus 5.5 (max) | GPT-6 Astra (max) |
|---|---|---|---|---|
| SWE-Marathon (long software projects) | 57.5 (xhigh) | 53.1 | 46.3 | 42.5 |
| DeepSWE 1.1 (long-horizon coding) | 79.0 | 78.8 | 74.2 | 74.1 |
| Harvey LAB-AA v1.1 (legal work, strict pass) | 13.1 | 10.0 | 4.2 | 8.6 |
| AutomationBench-AA (business workflows) | 45.4 | 39.1 | 41.2 | 41.6 |
| ProgramBench (program from scratch) | 59.0 | 50.0 | 70.5 | 52.5 |
| Terminal-Bench 4.0 (terminal coding) | 57.1 | 52.0 | 59.6 | 59.1 |
| HealthBench Pro (health questions) | 62.3 | 60.5 | 65.6 | 65.3 |
Source: Meta research blog and model card, October 12, 2026; numbers per Meta.
Three caveats. Meta uses a third-party or provider-reported result where one exists and otherwise the highest of its own run, a self-reported score, or the official leaderboard score, and it calls its runs of competitor models best-effort. SWE-Marathon is not a controlled comparison: Full ran at xhigh effort, and the GPT-6 Astra and Claude Fable 5.1 results used Codex and Claude Code. No row reports confidence intervals.
Harvey LAB-AA v1.1 needs its definition. In Artificial Analysis’s version of Harvey’s legal agent benchmark, a panel of three AI judges grades each deliverable on 120 private tasks. A task earns credit for the share of judges that find every rubric criterion met, and any material hallucination against the source documents sets it to zero; 13.1% is the average of that strict score. When the version launched on October 8, Artificial Analysis found material hallucinations in more than 60% of otherwise passing results across the models it tested.
5. Safety and limitations
Under its Meta Superintelligence Scaling Framework, Meta does not rule out that Full, without mitigations, reaches the high-risk threshold for chemical and biological misuse; with safety training and system-level classifiers, it rates residual risk as moderate or lower in every domain. On Meta’s charts, Full mishandled personal information in 2.2% of cases, Wonder in 3.4%, GPT-6 Astra in 4.6%, and Claude Fable 5.1 in 6.8%; Opus 5.5 is not charted. The model card lists 3D and spatial reasoning as a known limitation and says the model is not intended for people under 18.
What independent testing shows
Artificial Analysis scored both models at max effort on launch day (Intelligence Index v4.3.2, October 12, 2026).
| Model | Intelligence Index | Cost per index task |
|---|---|---|
| Claude Opus 5.5 (max with fallback) | 58 | $5.98 |
| Muse Wonder Full (max) | 56 | Not sold per token |
| GPT-6 Astra (max) | 53 | $3.26 |
| Gemini 4 Argon (high) | 53 | $1.99 |
| Muse Wonder (max) | 52 | $4.52 |
| GPT-6.1 Sol (max) | 52 | $0.72 |
| Muse Spark 1.3 (max) | 48 | $1.60 |
Source: Artificial Analysis pages for Muse Wonder and Muse Wonder Full.
Wonder, the model you can call from your own code, ties GPT-6.1 Sol at more than six times its cost per index task. Both Muse models produced more output tokens than the median on this suite, 140 million for Full and 130 million for Wonder against 88 and 85 million; the count includes reasoning and agent steps, not only final answers. AutomationBench-AA and Terminal-Bench 4.0 feed both this index and Meta’s table, so the two are not fully separate checks.
What it means for everyday users
Most people will meet Muse Wonder through Muse, Meta’s personal agent, which runs on mobile, Mac, and WhatsApp, handles email, reservations, and shopping, and asks for approval before sending an email or making a purchase. Meta’s blog says Wonder will soon power Muse, starting with subscribers, along with artifact generation and tasks that need more intelligence; Zuckerberg’s post says subscribers can use it today.
Meta’s demos show the intended work: a running plan built in a browser around your preferences, gift ideas, a wedding seating chart, an insurance claim, and, per the blog, understanding a lease or financial and health information.
Two cautions. For health, legal, and money questions, check the answer against the source: on health questions Opus 5.5 and GPT-6 Astra score higher in Meta’s own table, and strict legal grading gives every model tested a low score. On memory, the launch materials describe what Wonder can recall but not how to view or delete what it remembers; Muse says conversations are not shared with Meta’s ad systems.
Who should try Muse Wonder
- Developers on Muse Spark: test Wonder on multi-file coding and multi-hour agent runs, where Meta’s results are strongest, and keep Spark for high-volume or latency-sensitive steps, as Meta’s documentation advises. Expect a higher bill and measure it on your own workload.
- Teams on Claude Opus 5.5: the token rates match and Wonder was cheaper per index task, but it scored lower; the Muse model that matches Opus 5.5 in Meta’s results is Full, which your application cannot call.
- Legal agent builders: on Harvey LAB-AA v1.1 even the API model, at 10.0%, is ahead of the competitors in Meta’s table; check how often its drafts contradict your own source documents.
- Muse Code subscribers on Power or Maximum: try Full on long software projects, with Contributor mode off if your code should not be used for training.
- Cost-sensitive users: GPT-6.1 Sol and Muse Spark 1.3 cost far less per index task.
- Not suitable: anyone under 18, and applications that depend on precise 3D or spatial reasoning.
Reported versus published
| Question | Status on October 13, 2026 |
|---|---|
| Muse Wonder in the Muse app | Zuckerberg: subscribers can use it today. Meta’s blog: “soon”, subscribers first |
| Muse app plan prices and which plans get Wonder | TechCrunch reported Power at $20 and Maximum at $100 a month in September 2026; Meta’s launch materials list neither |
| Harvey LAB-AA v1.1 scores for the Muse Wonder models | Meta cites Artificial Analysis; the public leaderboard on October 13 still names Grok 4.7 (9.4%) as the top score |
| Parameter count and architecture | Not published |
| Open weights | Not mentioned; Artificial Analysis lists both models as proprietary |
| Full safety report | “Available soon” |
| Output speed | Not yet measured by Artificial Analysis |
| Free access | Zuckerberg: “over time”, no date |