Muse Wonder benchmark / benchmarks is not a single score. Meta’s comparison covers professional agents, coding and general evaluations. Muse Wonder Full scores 79.0 on the long-horizon coding benchmark DeepSWE 1.1; Muse Wonder scores 78.8, compared with 74.2 for Claude Opus 5.5 and 74.1 for GPT-6 Astra in Meta’s table. On the separate Artificial Analysis Intelligence Index, Meta lists 56 for Wonder Full, 52 for Wonder, 58 for Opus 5.5 and 53 for GPT-6 Astra. These scores answer different questions and are not one shared ranking.
Checked October 13, 2026. Scores are from Meta’s model page and model card. Meta says some competitor results come from providers or third parties, and not all harnesses are the same.
Muse Wonder benchmark scores vs GPT and Claude
The table below selects three results from Meta’s official model page. Wonder and Wonder Full use max reasoning effort; the comparison columns are also labeled max. Whether a higher score is better depends on each benchmark’s metric. Percentages, partial rewards and index scores are not interchangeable.
| Evaluation (task) | Muse Wonder Full | Muse Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| DeepSWE 1.1 (long-horizon software engineering, pass@1) | 79.0 | 78.8 | 74.1 | 74.2 |
| HealthBench Pro (open-ended health questions) | 62.3 | 60.5 | 65.3 | 65.6 |
| Artificial Analysis Intelligence Index (composite index) | 56 | 52 | 53 | 58 |
The results vary by task: Wonder Full is ahead of these two models on DeepSWE 1.1, while Opus 5.5 scores higher on HealthBench Pro and the Artificial Analysis Index in Meta’s table. The full Meta comparison also includes Claude Fable 5.1. Sources: Meta’s official benchmark table and the Muse Wonder model card (PDF).
What do Muse Wonder benchmarks measure?
DeepSWE 1.1 tests long-horizon software engineering. Meta’s model card describes 113 tasks across 91 repositories and five languages, graded with functional and regression tests. Meta uses Datacurve third-party results for Muse Wonder, Wonder Full and Claude Opus 5.5; it uses an OpenAI-reported result for GPT-6 Astra and the Anthropic system card for Claude Fable 5.1. The same benchmark name therefore does not mean every score was rerun by Meta in one common environment.
The Artificial Analysis Intelligence Index is a separate composite of evaluations run by Artificial Analysis. It is not Meta’s internal Personal Superintelligence Index (PSI), a measure of personal preference, task cost or application completion rate. Index versions and leaderboards change, so include the version and date when citing a score; see Artificial Analysis’ Meta model leaderboard for its independent ratings.
Why shouldn’t you rely on one leaderboard?
- Metrics differ. HealthBench Pro uses rubric scoring, DeepSWE reports pass@1 and the Artificial Analysis Index is a composite. Ranking their raw numbers together would be misleading.
- Setups differ. Meta says the models in its comparison table use max reasoning effort, but some outside results retain their own agents, tools or harnesses.
- Sources are not all independent. The model card uses third-party or provider-reported scores where available. It also describes Meta’s internal runs on proprietary competitors as best-effort; prompts, tools and runtime may not be fully tuned.
- Coding scores do not guarantee a deliverable. The model card notes that performance can vary with prompts, tools, quantization, serving configuration and agent harness. Your repository, acceptance tests and retries also affect results.
How can you compare Muse Wonder, GPT and Claude for your work?
Build a small set of real tasks with explicit pass criteria. Keep inputs, tool permissions, context, output constraints and acceptance tests the same. Record first-pass success, retries, full task cost and elapsed time; compare accepted results rather than one answer or token prices alone. For API tests, confirm the same model variant, reasoning effort, service tier and billing setup.
Meta lists Muse Wonder API rates of $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Muse Wonder Full is listed with the Muse Code Power and Maximum subscriptions. The billing models differ, and benchmark scores do not include your retry costs. See our guide to Muse Wonder subscription plans, prices and usage limits.
Muse Wonder benchmark FAQ
Is Muse Wonder Full always better than GPT-6 Astra or Claude Opus 5.5?
The table does not establish a general winner. Wonder Full scores higher on DeepSWE 1.1, while Opus 5.5 scores higher on Meta’s HealthBench Pro and Artificial Analysis Index rows.
Are Meta’s benchmark results independent?
Not all scores were rerun independently by Meta. The model card identifies third-party, provider-reported or model-provider results where used, and says its internal runs on proprietary competitors were best-effort. Check each benchmark’s method and source before citing a specific number.
What do 56 and 52 in Meta’s table mean?
They are the Artificial Analysis Intelligence Index values Meta lists for Muse Wonder Full (56) and Muse Wonder (52). They are not the model card’s Personal Superintelligence Index or percentages.
How do I know whether these scores matter for my project?
Test the models you are considering on your own code or work samples, with the same tools and explicit acceptance criteria. Track total task cost, failures and retries. Public benchmarks help shortlist models but do not replace project-specific evaluation.
Sources and method: Meta’s Muse Wonder model and benchmark page, the official Muse Wonder model card and Meta’s announcement. This article summarizes public evaluations; Token King did not run an independent head-to-head benchmark of Muse Wonder, GPT and Claude.