Meta Announces Muse Wonder AI Models Claiming Benchmark Wins Over OpenAI and Anthropic
Meta's benchmark claims put a third company's numbers on the table next to OpenAI and Anthropic, but the HealthBench Pro shortfall shows the lead is benchmark-specific, not across the board.
Key Facts
- Meta announced the AI models Muse Wonder and Muse Wonder Full on October 12, 2026.
- On the DeepSWE 1.1 benchmark, Muse Wonder Full scored 79.0 and Muse Wonder scored 78.8, above GPT-6 Astra at 74.1 and Claude Opus 5.5 at 74.2.
- On HealthBench Pro, Muse Wonder Full scored 62.3, below GPT-6 Astra at 65.3 and Claude Opus 5.5 at 65.6.
- Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is restricted to upper-tier Muse Code plans.
- Meta lists API pricing for Muse Wonder at $4 per million input tokens, $0.20 for cached input, and $20.00 for output.
Reporting from 2 sources: ASCII.jp, GIGAZINE.
Meta announced two frontier AI models, Muse Wonder and Muse Wonder Full, on October 12, 2026. The company presented benchmark results that it says surpass OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 on several tests. On DeepSWE 1.1, which measures long-term software development capability, Muse Wonder Full scored 79.0 and Muse Wonder scored 78.8, above GPT-6 Astra at 74.1 and Claude Opus 5.5 at 74.2. On Harvey LAB-AA, a legal document task, Full scored 13.1 against GPT-6 Astra's 8.6. The results are not uniform: on the healthcare benchmark HealthBench Pro, Full scored 62.3, below GPT-6 Astra's 65.3 and Claude Opus 5.5's 65.6. Both models share two features, Active Memory for retaining information across conversations and Active Focus for working through difficult tasks. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is limited to upper-tier Muse Code plans.
Meta described Muse Wonder as a model that delivers frontier-level performance at high speed, and Muse Wonder Full as the full version of the model. On the Artificial Analysis Intelligence Index, which aggregates results across benchmark tests, Muse Wonder scored 52 and Muse Wonder Full scored 56.
Meta cites active memory and active focus as the key elements for making the models useful as personal and business agents. Active memory records important items about the user and searches for detailed information relevant to the task. Active focus continues work on difficult tasks by exploring alternatives and adapting to new information.
On safety metrics, Meta presented graphs showing both the rate of executing unauthorized code and the rate of mishandling personal information lower than Anthropic's and OpenAI's models. Muse Wonder Full is not offered through the API. Muse Wonder API pricing is $4 per 1 million input tokens, $0.20 for cached input, and $20.00 for output.
Meta says the models were trained over 14 months at its gigawatt-class data center Prometheus, following a revamp of its AI development setup.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.