Meta Muse Wonder: Leading Claude Opus 5.5 & GPT-6 on Seven Benchmarks
On the 14 rows of Meta’s launch scorecard, a Muse Wonder model holds the top score in seven. Claude Opus 5.5 leads six and GPT-6 Astra leads one. Claude Fable 5.1 leads none, though it sits 0.2 points behind Muse Wonder Full on Vals Finance Agent V2 (58.9 against 59.1). All comparisons use maximum reasoning effort. Meta released the figures on October 12, 2026, alongside two models available through the Meta Model API and Muse Code. Muse Wonder Full is limited to Muse Code subscribers, and Meta says Muse Wonder will later power experiences in its Muse assistant, with subscribers first. The announcement and model card are published below.
Scores by row
The largest gaps appear in legal work and business automation. On Harvey LAB-AA, an Artificial Analysis version of Harvey’s Legal Agent Benchmark, Muse Wonder Full scores 13.1 and Muse Wonder 10.0, against 8.6 for GPT-6 Astra, 6.4 for Fable 5.1 and 4.2 for Opus 5.5. The metric counts only tasks whose deliverable meets every criterion and passes hallucination checks, so the top score corresponds to roughly one task in eight. On AutomationBench-AA, which checks the end state of simulated business workflows, Full scores 45.4 and the next model, GPT-6 Astra, scores 41.6.
Coding results split by benchmark. On DeepSWE 1.1, a long-horizon software engineering test, Full scores 79.0 and Muse Wonder 78.8, ahead of Opus 5.5 at 74.2, GPT-6 Astra at 74.1 and Fable 5.1 at 67.4. On SWE-Marathon, Full scores 57.5 and Wonder 53.1, against 46.3 for Opus 5.5. ProgramBench, which asks an agent to rebuild a complete program, reverses the order. Opus 5.5 scores 70.5, followed by Fable 5.1 at 55.5 and Full at 59.0 and GPT-6 Astra at 52.5, with Wonder at 50.0. On Terminal-Bench 4.0, Opus 5.5 (59.6) and GPT-6 Astra (59.1) are ahead of Full (57.1).
Several other rows go to competitors. On HealthBench Pro, Opus 5.5 scores 65.6 and GPT-6 Astra 65.3, against 62.3 for Full. On OSWorld V2.1, the computer-use benchmark, the primary partial-credit metric favors Opus 5.5 (81.8) and Fable 5.1 (80.7) over Full (78.4). On strict binary completion Fable 5.1 leads at 43.1, and no binary figure exists for Opus 5.5. On the Artificial Analysis Intelligence Index, Opus 5.5 scores 58 and Full 56. JobBench is nearly level, with Opus 5.5 at 75.4 and Full at 75.3.
Meta’s announcement says the series achieves state-of-the-art results on key benchmarks for personal and professional agents, memory and coding. The table supports that for memory and for the rows above where a Muse model leads, and it shows Opus 5.5 ahead on two of the four coding rows.
How the figures were selected
The model card describes its score selection in some detail. Where a comparable third-party or provider-reported result exists, Meta uses it. Otherwise it uses the highest of its internally validated run, a self-reported score, or an official leaderboard score. Competitor runs that Meta generated itself are described as best-effort, with prompts and tools that may not be tuned for proprietary models. Model refusals receive zero credit and stay in the denominator.
The sources differ from row to row. Harvey LAB-AA, AutomationBench-AA, Terminal-Bench 4.0 and the Intelligence Index come from Artificial Analysis. Vals Finance Agent V2 comes from Vals AI, and GDP.pdf from Surge. DeepSWE uses Datacurve’s results for the Muse models and Opus 5.5, OpenAI’s reported figure for GPT-6 Astra, and Anthropic’s system card for Fable 5.1. On HealthBench Pro, Meta ran the tasks itself with a GPT-5.4 grader and took the Claude figures from Anthropic’s system cards.
SWE-Marathon carries a caveat in the card itself. Abundant AI is the source for every model, but the GPT-6 Astra and Fable 5.1 results were produced with Codex and Claude Code respectively, so the row is not a controlled common-harness comparison. Muse Wonder Full’s score was reported at xhigh reasoning effort. The Personal Superintelligence Index, where Full scores 59.1 against 55.2 for Opus 5.5, is an internal composite of seven evaluations, three of which (FocusBench, VBT and the composite itself) Meta built.
Memory and sustained work
Meta groups two features under the names Active Memory and Active Focus. MemoryBench, an internal test, uses conversations of 64K to 100K tokens containing several target facts, distractor conversations, and injected turns that conflict with or update those facts. An LLM judge scores answers against weighted rubrics. Muse Wonder scores 85.6 and Full 85.5, against 81.5 for Opus 5.5, 81.4 for GPT-6 Astra and 79.0 for Fable 5.1. The smaller model edges the larger one by 0.1, which is within what run-to-run variation could produce. The models have a 1M token context length at inference, so the test conversations fill a tenth of the window or less.
Active Focus is measured with FocusBench, another internal evaluation, which covers tasks in housing, travel, IT and finance and asks how persistently an agent works through blockers. The model card does not list a FocusBench score in its table. The external check on sustained work is SWE-Marathon, described above.
Meta also reports a generalization test. It plots Muse Spark 1.3 and Muse Wonder Full on in-distribution tasks (those targeted in post-training) against out-of-distribution tasks, at several reasoning efforts. Full sits above the trend line the earlier model sets. Meta attributes this to a larger base model, whose post-training gains carry over more effectively to unfamiliar tasks. The announcement presents this as one scatter plot, and the text gives no underlying numbers.
Safety evaluations
Under the Meta Superintelligence Scaling Framework, a model reaches the high risk threshold when its deployment could substantially contribute to a threat scenario Meta has identified as severe. Meta states that without mitigations it does not rule out Muse Wonder reaching that threshold in chemical and biological risk. It applied safety post-training and system-level classifiers for that domain. Meta’s evaluations put the unmitigated model at moderate or lower for cybersecurity and loss of control. With mitigations, it assesses all domains at moderate or lower, and says an internal governance process reviewed that assessment before the deployment decision.
On cyber capability, Muse Wonder Full scored 88.4% on CyberGym and 58.9% on ExploitBench. On ExploitGym it captured 92 flags. A verification pass using a GPT-cyber verifier, confirmed independently with Gemini 3.1 Pro, found that 21 used the intended vulnerability and 71 used unintended ones. ExploitGym’s published ranking counts only intended exploits, so 21 is the comparable figure. Attack success rates for adversarial refusal were 0.8% on CyberBench in both single and multi turn settings, 1.1% and 2.3% on the PolyGuard cyber subset, and 9.0% and 8.0% on RedCode. The card reports these for Full only.
On biology and chemistry, refusal rates were 98.7% in English, 98.7% multilingual and 98.2% under adversarial prompts on SecureBio’s BioTIER, and 100.0%, 100.0% and 99.8% on Meta’s internal chemistry benchmark. The dual-use capability tests returned an average of 60.7% on the SecureBio knowledge suite, 89.0% on LAB-Bench ProtocolQA, 95.6% on SeqQA with tools, 100.0% on the ABC Bench liquid handling task and 79.4% on Meta’s BioDesign Tools evaluation. Two ABC Bench tasks, Fragment Design and Screening Evasion, have no scores because of a high refusal rate, and the card says refusals were dropped before scoring. That treatment differs from the zero-credit rule stated for the capability benchmarks. On loss of control, the model scored 87.3% on MASK, which tests honesty under pressure, and 0% on Agentic Misalignment, where lower is better.
Stated limits and next steps
The model card lists several limitations. Errors in multi-step reasoning can occur in scenarios the training data represents poorly. Visual and spatial reasoning, especially in 3D, remains a known weakness. Not every language in the pre-training data was evaluated, and performance varies with quantization, serving configuration, prompts, tools and the agent harness. The models are not intended for people under 18. For agentic deployments, Meta recommends human-in-the-loop confirmation of irreversible actions and additional guardrails suited to the use case.
Teams evaluating the models can start with the API and Muse Code at dev.meta.ai, and the benchmark names in Meta’s table link to official leaderboards where they exist. Rows sourced from Artificial Analysis, Vals AI and Abundant AI can be checked against those sites directly. Because the SWE-Marathon and ProgramBench rows used different harnesses and the HealthBench Pro grader was chosen by Meta, a trial on an organization’s own tasks will show more than the scorecard. Meta says a full report with additional safety evaluations will be available soon.
