Meta Releases Muse Wonder and Wonder Full — Rivals to GPT-6 Astra and Claude Fable

Meta has released two new AI models, Muse Wonder and Muse Wonder Full, positioning them as rivals to OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1. On internal benchmarks, Wonder Full outperformed its main competitors, particularly in coding tasks and legal analysis.

Two models, two performance tiers

Meta released two versions: the base Muse Wonder, which scored 54.3 points on the Personal Superintelligence Index, and the flagship Wonder Full, which scored 59.1. Both models are available to Muse Code subscribers, while the base Wonder version is integrated into the Meta Model API for developers.

Meta's developers position the models as direct competitors to OpenAI's and Anthropic's flagship offerings. The comparison with Google's Gemini 4 Argon remains tentative, since the market hasn't yet fully explored that model's capabilities. All tests used the models' maximum reasoning setting.

Coding: a clear edge for Wonder Full

In software development tasks, Meta's model showed its widest lead over competitors:

• DeepSWE 1.1 (solving complex tasks in existing repositories): Wonder Full — 79 points, GPT-6 Astra — 74.1, Claude Opus 5.5 — 74.2 • SWE-Marathon (long-term work on large projects): Wonder Full — 57.5 points, GPT-6 Astra — 42.5, Claude Fable 5.1 — 45.6, Claude Opus 5.5 — 46.3

However, in other coding benchmarks, the advantage is smaller or absent entirely:

• Terminal-Bench 4.0 (command line): GPT-6 Astra and Claude Opus 5.5 outpaced Wonder Full (59.1 and 59.6 versus 57.1) • ProgramBench (automatic program repair): Claude Opus 5.5 is notably ahead (70.5 versus Wonder Full's 59)

Legal analysis and overall performance

On the Harvey LAB-AA benchmark, which evaluates legal task performance based on documents and detailed instructions, Wonder Full significantly outperformed its rivals:

• Harvey LAB-AA: Wonder Full — 13.1 points, GPT-6 Astra — 8.6, Claude Fable 5.1 — 6.4, Claude Opus 5.5 — 4.2

On the composite Personal Superintelligence Index, which combines seven benchmarks covering professional and everyday tasks, the results broke down as follows:

• Muse Wonder Full — 59.1 points • Claude Opus 5.5 — 55.2 points • GPT-6 Astra — 54.5 points • Muse Wonder — 54.3 points • Claude Fable 5.1 — 49.9 points

Plans for the Muse AI agent and Meta's strategy

Meta plans to use the Wonder architecture as the foundation for its Muse AI agent. The updated agent will first roll out to paying subscribers, after which the company intends to develop specialized versions for a mass audience. Specific rollout timelines and technical details haven't been disclosed.

The launch of these models is part of Meta's aggressive strategy to reclaim influence in AI. The corporation has invested billions of dollars in attracting talent and building out computing infrastructure, while simultaneously expanding its product portfolio — including the image-generation model Muse Image.