Meta joins the pack

Meta’s Muse Wonder models were evaluated on Hex’s DataBench benchmark for agentic analytics and achieved a best score of 45%, placing slightly above GPT-5.6 Sol but below GPT-6 models and making them the first non-OpenAI or Anthropic models to approach leading performance. The evaluation found that Muse Wonder is relatively expensive, less token-efficient, and inconsistent as effort increases, with its performance declining from 45% at xhigh effort to 34% at maximum effort, though the authors note that their testing harness may not yet be optimized for Meta’s API. Muse showed notable persistence and adaptability on difficult tasks, successfully revising its approach when data labels or assumptions proved misleading, reflecting Meta’s “Active Focus” training emphasis on working through blockers. However, many failures stemmed not from flawed analysis but from overly confident opening claims, delayed caveats, and additional unsupported conclusions introduced at higher effort levels. The benchmark also penalized Muse’s tendency at maximum effort to pause for confirmation before completing some tasks, a behavior that may be useful in real collaborative settings but conflicts with DataBench’s single-response format. The authors conclude that clearer communication, better-calibrated persistence, and steadier effort scaling could make Muse Wonder a strong competitor for analytical work.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.