Meta's Muse Wonder Full beats Claude Opus 5.5 on 8 of 14 Meta-published benchmarks, but only the weaker variant is sold through the API

Meta Superintelligence Labs (MSL) released Muse Wonder, a two-tier model line for coding and long-running agent tasks. Both tiers have a 1,048,576-token context window, according to VentureBeat. The Russian-language channel AI ML Big Data described the launch as Meta's return to the frontier. AI Post said the top variant beats Claude Opus 5.5 on 8 of 14 published benchmarks and leads GPT-6 Astra and Claude Fable 5.1 on long-running coding, legal work and business tasks. Those comparisons come from Meta's own tables.

A row-by-row reading of Meta's numbers by Pasquale Pillitteri found that Muse Wonder Full wins eight rows and loses six against Opus 5.5. The base Wonder model wins five and loses nine. Full's largest leads are on legal and agentic coding: 13.1 versus 4.2 on Harvey LAB-AA, 79.0 versus 74.2 on DeepSWE 1.1, and 57.5 versus 46.3 on SWE-Marathon. Opus 5.5 leads on HealthBench Pro (65.6 versus 62.3), on OSWorld computer use (81.8 versus 78.4) and on ProgramBench (70.5 versus 59.0), where the gap is widest.

Independent measurement places Full behind Anthropic's flagship model. On the Artificial Analysis Intelligence Index, Muse Wonder Full scores 56, tied with Claude Sonnet 5.5 and behind Opus 5.5 at 58, LeoDex reported. The 8-of-14 headline therefore depends on which benchmarks Meta chose to publish. The legal and long-horizon coding evaluations where Full leads most are also the evaluations with the fewest independent reproductions.

The distribution terms matter as much as the scores. Developers can call base Muse Wonder through the Meta Model API, the commercial endpoint Meta opened with Muse Spark, at the same $4 input and $20 output per million tokens as Opus 5.5. Muse Wonder Full is not in the API. It runs only inside Meta's Muse Code assistant, on the $50 per month Power Usage tier or the $200 per month Maximum Usage tier, VentureBeat reported. Neither tier has downloadable weights. Forkast reported that Meta published no research write-up with the release, so parameter count, architecture and training compute remain undisclosed.