By ByteBot
Meta shipped its first frontier-class AI model on October 12. Muse Wonder and Muse Wonder Full are real, competitive, and in some benchmarks the best models available right now. There is a catch: the version developers can actually call via API is the weaker one — scoring 52 on the Artificial Analysis Intelligence Index against Claude Opus 5.5’s 58 — at the exact same $4 input / $20 output per million token price. Read the footnotes before you migrate anything.
Two Models, One Access Tier
Meta Superintelligence Labs, led by Alexandr Wang, launched two versions simultaneously:
- Muse Wonder — available via the Meta Model API, $4/$20 per million tokens, 1M token context window
- Muse Wonder Full — locked to Muse Code subscribers ($50+/month), no API access
Wonder Full is the model behind Meta’s headline benchmark numbers. It scores 56 on the Artificial Analysis Intelligence Index, within striking distance of Opus 5.5. The API model scores 52 — competitive with Fable 5.1 and GPT-6 Astra, but not ahead of them. Neither model offers downloadable weights.
What the Benchmarks Actually Say
Meta published a 14-row benchmark table. Wonder Full wins 6 rows. Opus 5.5 wins 6. The API version of Wonder wins one — MemoryBench, by 0.1 points.
| Muse Wonder (API) | Claude Opus 5.5 | GPT-6 Astra | |
|---|---|---|---|
| Price (input/output per MTok) | $4 / $20 | $4 / $20 | $10 / $50 |
| AA Intelligence Index | 52 | 58 | 53 |
| MemoryBench | 85.6 | 81.5 | 81.4 |
| DeepSWE 1.1 | 78.8 | 74.2 | 74.1 |
| SWE-Marathon | 53.1 | 46.3 | 42.5 |
| ProgramBench | 50.0 | 70.5 | 52.5 |
The story the table tells: the API version of Wonder is not better than Opus 5.5 overall. It is better at specific tasks — and those tasks happen to be ones developers care about.
Where Wonder Actually Wins
The places where even the API model leads are worth paying attention to. On SWE-Marathon — long-horizon software engineering tasks where agents work for extended periods — Wonder scores 53.1 against Opus 5.5’s 46.3. That’s a real gap. On DeepSWE, it scores 78.8 versus 74.2. On MemoryBench, it leads all models at 85.6.
These numbers line up with Meta’s marketed features. Active Memory persists information across sessions — not just within a single conversation window. Active Focus is designed to keep multi-step agents on task, recovering from errors and trying alternative approaches when one route fails. If your workload involves long-running coding agents or tasks that need to pick up context across sessions, those are legitimate reasons to test Wonder.
The official Meta announcement positions these models for “personal and professional agentic tasks” in law, finance, and health — domains where Wonder Full (not the API model) shows the biggest advantages.
The Infrastructure Gap
The model launched before the infrastructure was ready. Developers hit 429 rate-limit errors within minutes of the API going live. Meta has not published requests-per-minute or tokens-per-minute limits; the recommended workaround is exponential backoff. The Muse Code sandbox was broken on Mac and Linux at launch. In independent vision testing, Muse Wonder performed significantly worse than Opus 5.5 on image-to-HTML tasks, suggesting the vision stack did not receive the same optimization as the text and coding paths. The Hacker News launch thread captured the split reaction: respect for the benchmark numbers, frustration with the tooling.
The Privacy Note for Enterprise
Muse Code subscriptions — the only way to access Wonder Full — default to Contributor mode. That means Meta can use your submitted prompts, code, files, and outputs to train future models. API usage runs in Standard mode with no model training on submissions. For enterprise teams considering Wonder Full, that default needs to be understood and explicitly changed before any production workloads move over.
The Verdict
This is a real frontier debut from Meta. The company now has a model competitive at the top of the market, which matters regardless of whether it beats Opus 5.5 on the overall index. For developers: if your workloads are memory-heavy or involve long software engineering sessions, the API version of Wonder is worth evaluating — the MemoryBench and SWE-Marathon leads are genuine. For everything else, Opus 5.5 still outperforms the API model you can actually access. A detailed benchmark breakdown and VentureBeat’s analysis of the access gap are worth reading before committing.
