Meta released two models on 2026-10-12, Muse Wonder and Muse Wonder Full, and describes them as its most capable so far, with state-of-the-art results on personal and professional agentic tasks.[1] This Model Watch reads Meta's announcement, model card, model page and API documentation to answer the practical questions: what you can buy and at what price, what the specifications are, and how much weight the benchmark comparisons can bear.
Everything below comes from those four Meta pages, read on 2026-10-14. We did not run the models, so nothing here is our own testing. We also could not read the benchmark scores themselves: the announcement presents them in a chart, and the model card's text describes the evaluations without a score table for the capability benchmarks. We therefore quote no capability scores. Third-party summaries report numbers, but we did not open a primary source for them, so we leave them out.
Quick answer
- Two variants. Muse Wonder (
muse-wonder-1.0) is available through the Meta Model API and Muse Code. Muse Wonder Full (muse-wonder-full-1.0) is available to Muse Code subscribers.[1][3] - Price. $4.00 input, $0.20 cached input and $20.00 output per million tokens for Muse Wonder on the API.[3]
- Context. 1M tokens, listed as 1,048,576 in the API docs.[2][4]
- Inputs and outputs. Text, image, video and audio in; text out.[2]
- Status. The API documentation says Muse Wonder is in preview.[4]
- Benchmarks. Meta compares against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 and is open about several limits on those comparisons, covered below.[2]
Specifications and access
| Spec | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Model ID | muse-wonder-1.0 | muse-wonder-full-1.0 |
| Context window | 1M tokens | 1M tokens |
| Modalities | Text, image, video, audio in; text out | Same, per the model card |
| Where to get it | Meta Model API, Muse Code | Muse Code subscriptions only |
| Per-token API price | $4.00 in / $0.20 cached / $20.00 out per million | None listed |
| Subscription | Not applicable | Power Usage $50/month, Maximum Usage $200/month |
| Maximum output | Not listed on the model page | Not listed |
Two details deserve care. First, the model card treats the two names as one family ("Muse Wonder and Muse Wonder Full") and gives a single set of specifications, so we assume, but cannot confirm from the pages, that the architecture is shared and that the difference lies in serving configuration or capability. Second, Meta says that its safety risk assessment results come from Muse Wonder Full, "our most capable variant", which is why the assessment covers every model in the family.[2] That wording implies Full is the more capable of the two, though no page we read quantifies the gap.
The card lists the intended uses as general assistance, coding, personal and professional agents, domain knowledge work in areas such as law, finance and health, multimodal understanding and research support.[2] It states the model is not intended for people under 18.[2]
Two things we looked for and did not find: a model licence or a statement about releasing weights, and rate limits or data-retention terms for the API.[4] The card says use is subject to the Meta Model API Terms of Service, the Meta Model Acceptable Use Policy and the Muse Supplemental Terms of Service.[2] Read those before sending any sensitive data. If you are comparing models for cost, our API pricing comparison and cost-per-request explainer show how to turn per-token prices into per-task costs.
What the benchmark section of the model card actually says
Most launch posts show a table and leave the method to a footnote. Meta's model card puts the method in the main text, which lets us check it against the checklist in How to Read an AI Benchmark Claim. These are the statements that matter.
Reasoning effort and competitors
Muse Wonder results are produced at max reasoning effort, and the comparison models, GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, are also set to max reasoning effort.[2] Matching effort is the right default, because effort changes both score and cost. It does not make the cost equal: a table of scores at max effort says nothing about price per solved task. For how effort settings affect spend, see reasoning models, thinking tokens and cost.
The score selection rule
The card says that where a comparable third-party or provider-reported result exists, Meta uses it; otherwise it uses "the highest of our internally validated run, self-reported score, or official leaderboard score."[2] This is a transparent rule, and it is also a rule that picks the best available number for a model. Our reading: it can be a fair way to report a model's ceiling, and a poor way to estimate what you will see in your own harness, because it is not a median over repeated runs.
Controlled and uncontrolled comparisons
Meta separates two kinds of coding comparison. For controlled internal coding comparisons, every model uses the same benchmark-specific harness. External coding results keep the harness or agent product used by the reporting vendor, and the source is identified.[2] It then flags a specific case: on SWE-Marathon, the leaderboard results for GPT-6 Astra and Claude Fable 5.1 use Codex and Claude Code respectively, so "this row is not a controlled common-harness comparison", and Muse Wonder Full was reported at xhigh reasoning effort on that row.[2] A difference between models on that row mixes model quality, agent product and effort setting.
Best-effort competitor runs
For competitor runs Meta generates itself, the card says they use common settings where practical but "remain best-effort: prompts, tools, and runtime may not be optimally tuned for proprietary models."[2] That is an honest caveat. It also means a competitor's number in these rows can be lower than the vendor's own tuned setup would achieve.
Internal benchmarks and composites
Several evaluations in the card are Meta's own. The Personal Superintelligence Index is described as an internal composite of seven evaluations, and FocusBench, MemoryBench and the Verifiable Browsing Task are internal.[2] Internal benchmarks are not inherently suspect, but outsiders cannot inspect the tasks, so they cannot be reproduced. Independent leaderboards and third-party results are cited for others, for example Artificial Analysis for Harvey LAB-AA and AutomationBench-AA, Vals AI for Vals Finance Agent V2 and Datacurve for DeepSWE v1.1.[2]
Grading choices
The card documents grading details that change what a score means. Model refusals receive zero credit and stay in the denominator.[2] HealthBench Pro uses a GPT-5.4 grader against physician-written rubrics, and for the Claude models Meta uses the results Anthropic reports in its system cards.[2] OSWorld V2.1 reports a mean partial reward as the primary metric, and the card notes that Anthropic's Opus 5.5 figure there is a provider-reported partial score with no binary figure available.[2] Partial-credit and strict-pass metrics answer different questions, so check which one a row uses before comparing.
A decision guide
Based on the published pages, and as our interpretation rather than a test result:
- Shortlist it if you build long-running agents, coding agents or computer-use workflows, and want a fourth option beside the models in our Claude tier comparison. The context window and the multimodal inputs, including audio and video, are documented.
- Wait or test first if you need a stable, versioned production dependency. The docs call the model a preview, and we found no published rate limits.[4]
- Check the terms if your data is regulated or confidential. Retention and licensing terms are in separate legal documents we did not read.
- Do not choose on the headline table. Build a small eval from your own tasks. How to build LLM evals for your application walks through it.
What stays uncertain
- Capability scores: not in the text we could read.
- Whether Muse Wonder Full differs from Muse Wonder in architecture or only in serving and compute.
- Maximum output length, rate limits, data retention and licence.
- How the models behave on your workload, which no vendor table answers.
We will update this article with a dated changelog line when Meta publishes more detail.
Safety evaluation, briefly
The card says that, without mitigations, Meta does not rule out Muse Wonder Full reaching its "high" risk threshold in the Chemical and Biological domain, so it applies safety post-training and system-level classifiers. It rates the unmitigated model "moderate or lower" for cybersecurity and loss of control, and reports "moderate or lower" across all domains with mitigations in place.[2] These are Meta's own assessments, run under Meta's own framework, with some third-party-developed benchmarks. The card also recommends human-in-the-loop confirmation for irreversible actions in agentic deployments.[2] That recommendation applies equally to any agent you build; see How AI Agents Work, and How to Secure Them.
Limitations Meta lists
Meta's own limitations section says the model can produce inaccurate or biased responses, may err in multi-step reasoning on novel scenarios, remains limited in visual and spatial reasoning (especially 3D tasks), has not been evaluated on all languages in pre-training data, and varies with quantization, serving configuration, prompts, tools and agent harness.[2] The last point is the one to remember when reading any leaderboard row: the same weights can score differently under a different harness.
