Muse Wonder
Meta's first frontier models, built around Active Memory — context that survives across sessions.
Meta’s first frontier-grade models. The distinguishing feature is Active Memory: the model is trained to carry goals, constraints and earlier decisions across separate conversations, instead of leaving the developer to bolt a vector store onto the side.
What you build with it
- Long-running agent work where context has to survive between sessions
- Coding assistants that remember architectural decisions made last week
- Anything where re-explaining the project every session is the real friction
- Benchmarking your own memory approach against a trained one
What it gets you
- Memory is a trained capability with a published score, not a workaround
- Strong numbers on long-horizon coding tasks
- Priced in line with GPT-6 Sol and Claude Opus 5.5
- Models
- Muse Wonder · Wonder Full
- API price
- $4 in / $20 out per 1M
- Wonder Full
- Muse Code subscribers only
- MemoryBench
- 85.6 (Meta’s own)
- DeepSWE 1.1
- 79.0 — Wonder Full
Straight from the source
Watch instead of reading
The fast version, if you do not have time for the whole write-up.
Why it matters
Memory is the part of agent engineering everybody hand-rolls and nobody is happy with. The usual answer is a vector store bolted onto the side and a prompt that begs the model to use it. Meta trained the behaviour in instead, and published numbers for it — which at minimum gives the rest of us something to measure against.
Active Memory is the headline
Meta describes Active Memory as carrying information across conversations and pulling it back when relevant — keeping goals, constraints and earlier decisions in view as work progresses.
On Meta's internal MemoryBench, Muse Wonder scores 85.6 and Wonder Full 85.5, against Claude Opus 5.5 at 81.5, GPT-6 Astra at 81.4 and Claude Fable 5.1 at 79.0.
Those are the vendor's own numbers on the vendor's own benchmark, which is the weakest kind of evidence. But the framing is the useful part: memory as a trained capability with a score, rather than a retrieval hack you hope works.
The benchmark disagreement is the real lesson
On the long-horizon DeepSWE 1.1 coding test, Meta reports Wonder Full at 79.0 against 74.2 for Opus 5.5, 74.1 for GPT-6 Astra and 67.4 for Fable 5.1.
The third-party Artificial Analysis Index ranks Wonder Full at 56 — behind Claude Opus 5.5 at 58.
Both can be true. They measure different things, and scores from different benchmarks should never be compared directly. The practical takeaway for anyone choosing a model: vendor benchmarks tell you what the vendor optimised for, and your own eval on your own task is the only number that decides anything.
The good one is not on the API
Muse Wonder is available through the Meta Model API and Muse Code. Wonder Full — the stronger model — is reserved for Muse Code subscribers.
API pricing is $4 per million input tokens and $20 per million output, which puts it roughly level with GPT-6 Sol and Claude Opus 5.5.
Meta also calls out Active Focus: training for sustained execution, so the model tries another approach when a task gets hard, asks for clarification, and checks in before important actions.
Reach for it when
- Long-horizon agent work where context has to survive across sessions
- You are evaluating memory approaches and want a baseline to measure against
- You already work inside Muse Code and can reach Wonder Full
Watch out for
- Headline numbers are Meta's own, and an independent index ranks it lower
- Wonder Full is not reachable from the API — only standard Wonder is
- Never compare scores across different benchmarks; run your own eval