Meta's Muse Wonder Models Rival OpenAI and Google
Meta launched its closed Muse Wonder models, ranking fifth on the Vals Index while cutting per-task run costs below rivals from Google and Anthropic.
Independent testing by Vals AI places Meta’s closed model fifth in knowledge work benchmarks, with lower per-task run costs.
Meta introduced its Muse Wonder model family on October 12, 2026. The launch marks a notable strategic pivot: after years of centering its public AI strategy on open-weight Llama releases, the company chose a proprietary rollout, available strictly through commercial APIs and subscription tiers.
Early independent tests suggest the shift delivered immediate technical gains. In evaluations released by Vals AI, a platform measuring the economic impact of frontier systems against actual U.S. GDP sectoral weights, Muse Wonder Full secured fifth place on the Vals Index with a 64.57% accuracy score. The benchmark evaluates tasks across finance, legal analysis, software development, and tax accounting. The score placed Meta ahead of OpenAI’s GPT-6 Astra (63.13%) and just behind Anthropic’s Claude Sonnet 5.5 (67.04%) and Google’s top-ranked Gemini 4 Argon (68.90%). The standard Muse Wonder model placed tenth overall at 61.06%.
How Muse Wonder performed on the Vals Index
Performance diverged sharply across technical specialties. In CyberBench v1.1, an evaluation focused on cybersecurity workflows, the standard Muse Wonder model claimed first place with an 80.36% score.
Muse Wonder Full took third place on the Public Benefits Bench with 73.75%, and placed fourth on Harvey’s Legal Agent Benchmark with 21.67% on complex legal tasks. Speaking on The Information’s TITV program, Vals AI founder and CEO Rayan Krishnan stated that this release marks the first time Meta has delivered models competing directly at the frontier of intellectual work. Krishnan singled out the model’s analytical rigor and its ability to decline unworkable prompts instead of hallucinating.
Run costs and technical specifications
Operational pricing proved to be a clear differentiator during testing. Vals AI measured the average run cost for Muse Wonder Full at $8.56. Running the identical test battery cost $15.68 on Gemini 4 Argon and $21.34 on Claude Sonnet 5.5.
On Meta’s commercial API, pricing stands at $4.00 per million input tokens and $20.00 per million output tokens. The models support a 1-million-token context window with maximum output generations of up to 131,000 tokens per API call. Meta trained the models at its Prometheus data center in New Albany, Ohio.
API rate limits and long-term enterprise use
Despite solid benchmark numbers, practical rollout challenges surfaced right away. Developers reported hitting strict rate limits within hours of the API going live.
Beyond initial server constraints, Meta gave no indication that it intends to release open weights for the Muse family, creating friction for teams accustomed to Llama’s hosting flexibility. The architecture’s true reliability across continuous corporate workflows will depend on how it performs over months of production deployment.