Meta returns to the race for the most powerful AI models with Muse Wonder and Muse Wonder Full. Launched on October 12, the two models feature a context window of one million tokens. On sever
Meta returns to the race for the most powerful AI models with Muse Wonder and Muse Wonder Full. Launched on October 12, the two models feature a context window of one million tokens. On several benchmarks published by Meta, Wonder Full outperforms GPT-6 Astra and Claude Opus 5.5. Independent tests are more nuanced: Anthropic remains ahead in the overall ranking.
Meta presents Muse Wonder as an important step toward “personal superintelligence” advocated for several months by Mark Zuckerberg. The launch comes after a much less comfortable period for the group. In July, Zuckerberg still acknowledged that Meta was progressing too slowly in AI, despite considerable investments and a major reorganization of its teams.
Muse Wonder changes the picture somewhat. The model has a context window of one million tokens, enough to handle very large amounts of text during a single task. Meta also introduces two core functions: Active Memory and Active Focus.
Active Memory allows the AI to retain some useful information across different conversations. Active Focus is intended to allow it to stay longer on a complex problem, try multiple approaches, and continue its work when a first method fails.
The goal therefore goes beyond simple response generation. Meta wants agents capable of working over multiple stages on code, documents, or professional tasks. Muse Wonder is available in the Meta Model API and Muse Code. The Wonder Full version is currently reserved for Muse Code subscribers.
Start your crypto adventure safely with OKXThis link uses an affiliate program.Meta supports this release with several particularly high results. On its Personal Superintelligence Index, Wonder Full scores 59.1 points. Claude Opus 5.5 reaches 55.2 points, and GPT-6 Astra 54.5.
The gap increases on certain specific tasks. SWE-Marathon measures an AI’s capacity to work for a long time on programming problems. Wonder Full scores 57.5, versus 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra.
Meta’s model also dominates on the Legal Agent Benchmark used to test professional legal tasks. Its score reaches 13.1, against 8.6 for Astra and 4.2 for Opus 5.5. Memory is another strong point. On MemoryBench, Wonder and Wonder Full score around 85.5. Astra reaches 81.4.
Meta now has a model much better suited to Muse, its AI agent capable of browsing the web, filling forms, and making purchases. One important detail remains: these early figures come from evaluations published by Meta. They do not mean that Wonder Full is automatically better than Astra in all areas.
Artificial Analysis provides another point of comparison. Its Intelligence Index assigns 56 points to Muse Wonder Full and 52 to Muse Wonder. The progress remains significant. Muse Spark 1.3 had only scored 48 points six weeks earlier.
Wonder Full thus enters the group of the most powerful AI models currently available. Anthropic still holds the advantage on this independent evaluation. Claude Opus 5.5 achieves a score higher than Meta’s model.
Wonder also shows some weaknesses. Artificial Analysis notably notes a lag in general knowledge tests. Wonder Full scores 30 on AA-Omniscience, versus 43 for GPT-6 Astra and 46 for Claude Opus 5.5.
The model stands out more on agents, scientific reasoning, and certain programming tasks. Price adds another comparison element. Meta charges 4 dollars per million input tokens and 20 dollars per output token for Muse Wonder. GPT-6 Astra costs 10 dollars per input and 50 dollars per output token.
Meta does not win all benchmarks. But the gap has narrowed considerably. A few weeks after the release of GPT-6 Astra, presented by OpenAI as a major new step in its AI, Zuckerberg now has a model capable of competing directly in several categories. After months of chasing OpenAI and Anthropic, Meta finally returns to the top group.
