Meta released Muse Wonder and Muse Wonder Full on Monday, October 12, and it is calling them its first frontier AI models, meaning the same tier as the best systems from OpenAI and Anthropic. According to TechCrunch, both come out of Meta’s Superintelligence Lab and will power the Muse agent.
What Meta shipped
Muse Wonder is larger and more capable than Muse Spark, the lightweight model Meta released earlier and built for low latency. In Meta’s own testing, the new models land close to Fable 5.1 and GPT-6 Astra, and the company says Wonder beats those rivals on MemoryBench, a memory test.
Memory is the main pitch. Meta calls it “Active Memory”: the model keeps and combines what you tell it over days, weeks and months. There is also “Active Focus,” which makes the model stick with hard tasks, try a different approach when it gets stuck, and ask follow-up questions. Both models take text, image, video and audio input and have a one-million-token context window (how much text the model can consider at once).
How the independent benchmarks look
Artificial Analysis, as reported by The Decoder, scores Wonder Full at 56 and Wonder at 52 on its Intelligence Index, up from 48 for Muse Spark 1.3. Meta says Wonder Full ranks in the top three on six of ten benchmarks.
The biggest gains are in agentic work, where the AI carries out steps on its own:
- On the Coding Agents Index, Wonder Full scores 68, tying Claude Code running Sonnet 5.5. Wonder rose from 54 to 63.
- On Terminal-Bench, Wonder Full went from 33% to 57% while using 30% fewer output tokens.
- On ProofBench, a private math-proof test, it solves 94 of 100 problems, up from 58. The top models solve 99 or 100.
- Its hallucination rate (made-up answers) fell from 33% to 23%.
Not everything improved. Finance and legal tasks stayed at the predecessor’s level, down by up to two points. On the AA-Omniscience knowledge test, Wonder Full scores 30, well behind Claude Opus 5.5 (46) and GPT-6 Astra (43), which suggests a smaller base model.
Price and availability
Through Meta’s API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens, compared with $1.25 and $4.25 for Muse Spark. On the Artificial Analysis index, the cost per task jumped from $1.60 to $4.52, nearly triple. For comparison, the same measure puts Claude Opus 5.5 at $5.98 and GPT-6 Astra at $3.26.
Wonder is available through the API and Muse Code, and will soon power the Muse app, starting with paying subscribers. Wonder Full is limited to Muse Code users for now and is not in the API, with no price announced yet. Meta trained the models at its Prometheus data center in Ohio, and says it plans to build specialized models from Wonder for all users over time.
Why it matters
This is Meta’s first frontier release since Alexandr Wang restructured its AI work in August 2025. Since then the company shipped Muse Spark and the Muse agent but had nothing at the level of the leaders. If you want background on how these systems work, our explainer on what a language model is covers why they sometimes make things up.
For a look at how a rival handles its own agents, see also the story of Anthropic Claude sending a false tip to Philadelphia police.
