Muse Wonder is Meta's AI model, released 12/10/2026. Meta's model card describes Muse Wonder as a model built for personal and professional agentic tasks, evaluated at max reasoning effort against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. It leads on Harvey LAB-AA, DeepSWE 1.1 and SWE-Marathon but trails Opus 5.5 and Astra on several rows such as ProgramBench, Terminal-Bench 4.0 and HealthBench Pro. Safety risk assessments were run on Muse Wonder Full and apply to the family.
Visit Muse WonderAI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does Meta Muse Wonder score on independent benchmarks?
Meta Muse Wonder ranks #12 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 43.6, as measured on 22 November 2026.
Where does Meta Muse Wonder rank against other models?
How good is Meta Muse Wonder at agentic coding?
How does Meta Muse Wonder do on individual tests?
| Evaluation (run by Artificial Analysis) | Meta Muse Wonder | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 52% | 67.2% |
| FrontierCode 2.0 | 27.9% | 48.8% |
| Humanity's Last Exam | 50.1% | 74.3% |
| GDPval-AA (Elo) | 1770 | 1866 |
How has Meta Muse Wonder's score changed over time?
We have tracked Meta Muse Wonder on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the Meta Muse Wonder system card report?
Meta's model card describes Muse Wonder as a model built for personal and professional agentic tasks, evaluated at max reasoning effort against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. It leads on Harvey LAB-AA, DeepSWE 1.1 and SWE-Marathon but trails Opus 5.5 and Astra on several rows such as ProgramBench, Terminal-Bench 4.0 and HealthBench Pro. Safety risk assessments were run on Muse Wonder Full and apply to the family.
These are Meta's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Sources: Muse Wonder Model Card (12/10/2026), Introducing Muse Wonder models (12/10/2026), Muse Wonder research blog (12/10/2026). Last checked 22/11/2026.
Agents & tool use
| Benchmark | Meta Muse Wonder | Muse Wonder Full | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| Personal Superintelligence Indexinternal composite of seven personal and knowledge-work evaluations; max | 54.3 points | 59.1 | 49.9 | 54.5 |
| Harvey LAB-AAlegal agent work; Artificial Analysis v1.1 with hallucination checks; max | 10% | 13.1% | 6.4% | 8.6% |
| Vals Finance Agent V2private test set accuracy; Vals AI; max | 56.6% | 59.1% | 58.9% | 53.5% |
| AutomationBench-AAstrict pass@1; Artificial Analysis; max | 39.1% | 45.4% | 32.1% | 41.6% |
| JobBench65 professional tasks, mean reward; max | 70.5% | 75.3% | 73.1% | 66.1% |
| OSWorld V2.1 (partial credit)108 workflows, mean per-task partial reward; max | 72.6% | 78.4% | 80.7% | 75.8% |
| OSWorld V2.1 (binary)strict full-completion rate; max; no binary figure for Opus 5.5 | 34.3% | 39.4% | 43.1% | 34.3% |
Knowledge & science
| Benchmark | Meta Muse Wonder | Muse Wonder Full | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| HealthBench Pro525 questions, no tools, length-normalised rubric score; max | 60.5% | 62.3% | 62.1% | 65.3% |
Multimodal
| Benchmark | Meta Muse Wonder | Muse Wonder Full | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| GDP.pdfprofessional work grounded in PDFs; Surge; max | 30.8% | 33.2% | 27.6% | 34.2% |
Long context
| Benchmark | Meta Muse Wonder | Muse Wonder Full | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| MemoryBenchinternal recall test over 64K to 100K token conversations; max | 85.6% | 85.5% | 79% | 81.4% |
Coding
| Benchmark | Meta Muse Wonder | Muse Wonder Full | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0-AAArtificial Analysis accuracy; max | 52% | 57.1% | 52% | 59.1% |
| DeepSWE 1.1pass@1; Datacurve third-party results; max | 78.8% | 79% | 67.4% | 74.1% |
| SWE-Marathonultra-long-horizon engineering, pass@1; Abundant AI; native harnesses so not a controlled comparison; xhigh reasoning effort for Muse Wonder Full | 53.1% | 57.5% | 45.6% | 42.5% |
| ProgramBenchshare of rollouts passing at least 95% of hidden tests; mini-swe-agent; max | 50% | 59% | 55.5% | 52.5% |
Other
| Benchmark | Meta Muse Wonder | Muse Wonder Full | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3Artificial Analysis composite; max | 52 points | 56 | 53 | 53 |
What did safety testing find for Meta Muse Wonder?
Safety level: Meta Superintelligence Scaling Framework: 'moderate or lower' across Cybersecurity, Chemical and Biological, and Loss of Control with mitigations in place; the high-risk threshold in Chemical and Biological cannot be ruled out without mitigations
- Without mitigations Meta does not rule out the model reaching the high-risk threshold in Chemical and Biological, so it applies safety post-training and system-level classifiers.
- Unmitigated, the model is assessed at 'moderate or lower' for Cybersecurity and Loss of Control; with mitigations all domains are 'moderate or lower'.
- Results are from Muse Wonder Full, the most capable variant, and the assessment covers every model in the family.
- Cyber robustness (attack success rate, lower is better): CyberBench 0.8% single and multi turn; PolyGuard cyber 1.1% single and 2.3% multi turn; RedCode 9.0% single and 8.0% multi turn.
- CBRN capability proxies: SecureBio VCT 55.9%, HPCT 66.5%, MBCT 59.6%; LAB-Bench ProtocolQA 89.0% and SeqQA with tools 95.6%; two ABC Bench tasks could not be scored because of a high refusal rate.
- Loss of control: MASK honesty 87.3% and Agentic Misalignment harmful action rate 0%. Known limitations include weak 3D spatial reasoning and the model not being intended for under-18s.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our Meta Muse Wonder deep dive.
Meta Model Timeline
1M tokens input; up to 512K output tokens context
1M tokens at inference context
1M tokens at inference context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
