Meta Superintelligence Labs released two frontier models on Monday 12 October. Muse Wonder is on the Meta Model API at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full, the stronger of the pair, cannot be called through the API at all. It is only available inside Muse Code, Meta’s terminal coding agent, on plans starting at $50 a month. If you build on APIs, the model you can integrate scores 52 on the Artificial Analysis Intelligence Index, not 56.
What happened
TechCrunch describes the launch as Meta’s first attempt at a frontier model. It is also the first frontier release since Alexandr Wang, the Scale AI founder, restructured Meta’s AI efforts. The models were trained at Meta’s Prometheus data centre in central Ohio.
Meta’s earlier Muse Spark was a lightweight model built for low latency. The Wonder models are larger and aimed at coding and long-running agents. Meta’s two headline features are “Active Memory”, which carries context across sessions, and “Active Focus”, which is meant to keep multi-step agents from drifting.
Independent scoring puts the pair near the top of the market, but not at it. AI Weekly’s summary of the Artificial Analysis results has Wonder Full level with Claude Sonnet 5.5 and behind Claude Opus 5.5:
| Model | AA Intelligence Index | How you get it |
|---|---|---|
| Claude Opus 5.5 | 58 | Public API |
| Claude Sonnet 5.5 | 56 | Public API |
| Muse Wonder Full | 56 | Muse Code only ($50 or $200 a month) |
| GPT-6 Astra | 53 | Public API |
| Muse Wonder | 52 | Meta Model API, $4/$20 per M tokens |
Meta also published its own benchmark chart. Fello AI’s breakdown of that scorecard counts Wonder Full winning 7 of 14 rows. Its biggest leads are on SWE-Marathon (57.5 against Opus 5.5’s 46.3) and DeepSWE 1.1 (79.0 against 74.2).
Two of those seven wins are on benchmarks Meta built itself. The model card also calls Meta’s runs of rival models “best-effort” and says they may not be optimally tuned. None of that makes the numbers wrong, but you should treat the margins as a ceiling rather than a floor.
What it means for builders
The API pricing is less flattering than the benchmarks. Fello points out that Muse Wonder’s $4 input, $0.20 cached and $20 output per million tokens is identical to Opus 5.5 on every line. That is despite Wang telling CNBC it would be priced “meaningfully below” OpenAI and Anthropic. Sonnet 5.5 scores 56 at half that price.
So for most API workloads there is no price argument for switching. The case for testing it is narrower: long coding projects and legal or document-heavy agent work, where Meta’s margins are widest.
The more important issue is access. AI Weekly’s editor’s note says Meta is trading developer-API reach for consumer monetisation. It adds that enterprise buyers looking for a frontier endpoint outside OpenAI and Anthropic only get the smaller Wonder.
Our read is blunter. A $50 a month gate is not a pricing tier; it is a platform. Every rival model in the table above is sold by the token through a public API, and Wonder Full is not.
That has a practical cost. Suppose a team comes to rely on Wonder Full inside Muse Code for real work, such as long refactors and multi-file changes. There is no endpoint to point a pipeline at and no way to put it behind your own abstraction layer. There is no reseller to fall back on either.
The Muse Wonder you can reach through OpenRouter, or through AI/ML API, which lists it at $5.5016 input and $27.508 output per million tokens, is a different and weaker model. Fello also notes that neither the launch post nor the model card mentions open weights.
Nothing Meta has published says the Muse Code exclusivity will change. Our guess, and it is only a guess, is that a future Wonder Full will follow the same pattern. Plan on that basis rather than hoping for an API release.
What to do now
- Test before you switch. Benchmark Muse Wonder on the API against Opus 5.5 and Sonnet 5.5 using your own tasks, not Meta’s chart. Weight the comparison towards long coding and document-heavy work if that is where you hoped it would help. The Meta Model API works with OpenAI-SDK-compatible clients, so a trial is cheap to set up.
- Keep Wonder Full out of production. Treat Wonder Full in Muse Code as a developer tool, like an editor subscription, not as a production dependency. Anything customer-facing or scheduled should run on a model you can call, version and swap.
- Check where lock-in has already crept in. Look for places where subscription-only tools are carrying real workflow load without anyone deciding they should. An AI readiness audit of your architecture is a sensible place to surface that.
This is the same trade-off as any custom-versus-off-the-shelf decision. Convenience today is paid for in exit options later. With Wonder Full, the exit option is not there to begin with, so make that choice knowingly.
Frequently Asked Questions
Can I use Muse Wonder Full through an API?
No. Muse Wonder Full is only available inside Muse Code on the $50 Power Usage and $200 Maximum Usage plans. Only the smaller Muse Wonder is served through the Meta Model API and third parties such as OpenRouter.
Is Muse Wonder cheaper than Claude Opus 5.5?
No. According to Fello AI, Muse Wonder's API price of $4 input, $0.20 cached and $20 output per million tokens is identical to Opus 5.5. Opus 5.5 also scores higher on the Artificial Analysis Intelligence Index, 58 against 52.
Are Muse Wonder's weights open?
No. Fello AI notes that neither Meta's launch post nor the model card mentions releasing weights for either Muse Wonder model.
Want this working in your business?
D3X Solutions LABS builds AI agents and automation that pay back fast. Set up a free 30-minute call and we will map your first win together.
Get in touch
