Meta has released Muse Wonder and Muse Wonder Full, its latest models for coding and agents that work through longer tasks. Wonder is available in preview through the Meta Model API as muse-wonder-1.0 and in Muse Code. Full, the stronger of the two, is restricted to Muse Code’s Power Usage and Maximum Usage subscriptions.
The coding results are the standout. In Meta’s published comparisons, both models beat GPT-6 Astra and Claude Opus 5.5 on DeepSWE and SWE-Marathon, two evaluations of extended software-engineering work.
Where Wonder pulls ahead
On DeepSWE 1.1, Wonder Full scores 79.0% and Wonder 78.8%, against 74.2% for Opus 5.5 and 74.1% for Astra. All four are labelled at maximum reasoning effort in Meta’s chart. The two Wonder variants are almost level here, despite their different access requirements.
The gap between them is larger on SWE-Marathon: Full reaches 57.5 at xHigh effort, while Wonder records 53.1 at max. Opus 5.5 scores 46.3 and Astra 42.5 at max in the same table.
That lead does not carry across every test. In the Artificial Analysis Intelligence Index chart reproduced by Meta, Wonder Full scores 56, above Astra’s 53 but below Opus 5.5’s 58. Standard Wonder scores 52. Meta’s wider table also puts Opus ahead on Terminal-Bench 4.0-AA and ProgramBench. Long-running coding is the clearest strength in Meta’s results.
Remembering context and sticking with a task
Meta groups the agent improvements under two names. It says Active Memory retrieves relevant details from earlier conversations. Active Focus describes working through blockers, changing approach and asking for clarification when needed. Both models have a roughly one-million-token context window, according to the model page.
Pricing and access
Meta’s API pricing lists the following rates for Muse Wonder 1.0. There is no extra long-context charge as the context window fills up.
- Input
- $4Uncached input tokens.
- Cached input
- $0.20Input tokens served from the prompt cache.
- Output
- $20Generated output tokens.
Full has no listed API rate. Meta offers it through the $50-a-month Power Usage and $200-a-month Maximum Usage Muse Code plans. Wonder is also due to reach the Muse assistant, starting with subscribers; the API and coding-tool release is available now.
Sources & references
- Meta: Muse Wonder announcement and evaluation results ↗
12 October 2026. Launch details, Active Memory and Active Focus, and Meta’s published benchmark comparisons. Both charts in this article are from this announcement.
- Meta: Muse Wonder models and availability ↗
Context window, Wonder Full subscription access and launch artwork. Checked on 12 October 2026.
- Meta Model API: pricing and rate limits ↗
Standard-mode rates for muse-wonder-1.0, per million tokens. Checked on 12 October 2026.
- Lumina on X: Muse Wonder benchmark highlights ↗
Our launch coverage. The figures originate with Meta’s release; this is not a separate evaluation.
- Lumina on X: Muse Wonder API pricing ↗
Our pricing post, checked against Meta’s developer documentation.
