Muse Wonder Benchmarks: Meta Closes In but Opus 5.5 Still Leads

What Meta Shipped

Meta Platforms has released Muse Wonder, its first model family that credibly competes at the frontier. The lineup includes a standard Muse Wonder and a larger Muse Wonder Full. One account dates the launch to October 12, 2026 3. Bloomberg reports that the model was known internally as "Watermelon." It is the most capable system Meta has produced since Mark Zuckerberg rebuilt his AI organization through an expensive hiring campaign that put new leadership in charge 4. Meta says the model matches OpenAI's ChatGPT and Anthropic's Claude on coding and on financial and legal reasoning 4.

The independent numbers support a narrower claim. Meta is now close to the top, but it is not tied with the leader.

The Independent Scorecard

Artificial Analysis evaluated both models quickly. On its Intelligence Index, Muse Wonder Full scores 56 and Muse Wonder scores 52. Anthropic's Claude Opus 5.5 sits at 58 123. That puts Wonder Full level with Claude Sonnet 5.5 and ahead of OpenAI's GPT-6 Astra, Google's Gemini 4 Argon and Anthropic's Claude Fable 5.1, each of which scores 53 2. The standard Wonder lands one point below GPT-6 Astra 2.

Reports describe the index slightly differently. One says it combines nine evaluations 1. Another says ten, spanning reasoning, knowledge, coding and professional tasks 2. A third cites version 4.3.2 of the index 3. The detail matters less than the ranking, which every outlet reports the same way: Opus 5.5 first, Wonder Full second-tier alongside Sonnet 5.5.

Independent testing also turned up domains where Meta clearly leads. On the Harvey LAB-AA legal benchmark, Wonder Full scores 13.1 against 4.2 for Opus 5.5. On AutomationBench-AA, which measures business workflows, it posts 45.4 versus 41.2 2. These results line up with Meta's own claims about legal and professional work 4.

Reading Meta's Own Chart

Meta's published comparison looks better than the aggregate. Wonder Full wins 7 of the 14 rows against Fable 5.1, GPT-6 Astra and Opus 5.5 3. The biggest margins come on long-horizon coding. It scores 57.5 to 46.3 on SWE-Marathon and 79.0 to 74.2 on DeepSWE 1.1 13.

The caveats are significant. Two of those seven wins come on benchmarks Meta built itself 3. One is a memory evaluation where both Wonder models score around 85.5 13. Wonder Full also trails Opus 5.5 on ProgramBench, 59.0 to 70.5, and on Terminal-Bench 4.0 and HealthBench Pro 1. One critique calls the table carefully curated, emphasizing strengths and leaving out weaknesses 1. A vendor chart that wins half its rows, partly on home-grown tests, is a selective view.

Availability and Rough Edges

Access may matter more to developers than the scores. Only the weaker Muse Wonder is available through Meta's API, so the 56-scoring Full model that produced the headline results cannot be used there 2. A developer-focused writeup is blunter. It reports early users hitting HTTP 429 rate-limit errors, raises privacy concerns about the model reading private messages, and notes that prediction market Polymarket gives Meta about a 1% chance on a related wager 1. Those claims come from a single, openly skeptical source and should be treated as early impressions rather than settled facts.

Why It Matters

The distance Meta has covered is the real story. Roughly fourteen months ago it was seen as a mid-tier open-weight lab. It now trails the strongest independently ranked model by two points 1. That counts as progress for Zuckerberg's costly reorganization, and Bloomberg frames the launch as the first real test of whether that spending paid off 4.

The framing still needs correcting. Meta's leadership describes parity 4, but the independent data shows a competitive runner-up. Wonder Full is excellent in specific areas, especially legal agents, business automation and marathon coding tasks. It is behind Opus 5.5 on the broad composite measure and on several coding and health evaluations that Meta's own chart concedes 123.

My conclusion is that Meta has earned a place in frontier conversations without taking the lead. Developers choosing a model should rely on the independent index and test against their own workloads, not on vendor charts. They should also note that the version most deserving of attention is the one Meta has not yet opened to API customers 2. If Meta releases Wonder Full broadly and resolves the reported reliability problems 1, the two-point gap could become a real contest. Until then, Opus 5.5 remains the benchmark to beat.