Meta shipped two frontier models on 12 October and only one of them has an API. Developers get muse-wonder-1.0 at $4 in and $20 out per million tokens. The stronger sibling, Muse Wonder Full, lives inside Muse Code subscriptions and nowhere else. We read Meta's model card next to Artificial Analysis's independent run, because the launch charts mix the two models more than you'd guess.
The short answer
Muse Wonder 1.0 is on the Meta Model API with a 1M-token context, at $4 input, $0.20 cached input and $20 output per million tokens, plus $2.50 per 1,000 web searches. Rate limits start at 3,000 requests and 4 million tokens a minute. Artificial Analysis scores it 52 on its Intelligence Index v4.3 at max effort; Wonder Full scores 56, but you can't call Full from code. At the same list price, Claude Opus 5.5 scores 58. No weights for either model.
Two models, one API
The naming makes this easy to miss. Meta's announcement talks about "Muse Wonder and Muse Wonder Full" in the same breath, and the benchmark chart in the blog post sets them side by side. But the model page on dev.meta.ai is blunt about access: muse-wonder-1.0 is on the Model API and in Muse Code, while muse-wonder-full-1.0 is "available by subscription in Muse Code". That's it. There's no API price for Full because there's no API for Full.
Meta doesn't explain what separates them. The blog says Full is "more capable" and stops there. No parameter counts, no word on whether Full simply thinks longer or is a different network. So when you see a headline number, check which name sits next to it. Artificial Analysis gives the cleanest split we found: 56 for Wonder Full and 52 for Wonder, both at max reasoning effort, up from 48 for Muse Spark 1.3. Full ties Claude Sonnet 5.5. Wonder lands below GPT-6 Astra and Gemini 4 Argon (53 each).
The model card adds one more wrinkle that's easy to skim past. Its risk assessment section runs every safety evaluation on Muse Wonder Full, "our most capable variant", and says that's why it covers the whole family. Reasonable. It also means the only detailed evaluation page Meta published is mostly about the model API customers don't get.
What $4 and $20 buys you
Meta's pricing page lists Muse Wonder in Standard mode only. That matters, because Muse Spark 1.3 also has a Contributor tier at $0.10 in and $0.20 out where Meta can train on your prompts. Wonder has no such discount on the API, and Standard content isn't used for training. Meta also says it injects "a small amount of steering context" into every prompt and doesn't bill those tokens, which is fair, though it's another reason your own token counts won't match a raw tokenizer exactly.
Now compare. Claude Opus 5.5 sits at exactly $4, $20 and $0.20 cached on our LLM API pricing tracker, and it scores 58 on the same index. Same sticker, six points lower. GPT-6 Astra lists at $10 and $50, two and a half times Wonder on both legs. But here's the number we'd look at before the list price: Artificial Analysis puts Wonder's cost to run one index task at $4.52, against $3.26 for Astra and $1.99 for Gemini 4 Argon. Wonder is cheaper per token than Astra and dearer per task, because it chews through far more tokens at max effort. That's a single evaluation suite, and your workload won't look like it. Still, it's a warning against reading "$4 vs $10" as "less than half the bill".
There's a cheaper way to reach Full, if you don't need an API. Press coverage from VentureBeat and others puts Muse Code at $5 to $200 a month, with Full on the $50 and $200 tiers, and cites Meta's help center saying subscriptions default to Contributor mode, where Meta can train on your code. We couldn't open that help page ourselves, so treat the default as reported, and check the toggle before you point it at a private repo.
Read the evaluation footnotes first
Meta's model card is more candid than its chart. Every Muse result was run at max reasoning effort. Competitor runs done in-house are "best-effort", and Meta says the prompts and tools "may not be optimally tuned" for other vendors' models. The score-selection rule is the part we'd underline: where no third-party or provider-reported result exists, Meta uses "the highest of" its own validated run, a self-reported score or a leaderboard score. Taking the best of several numbers for each cell flatters whoever's picking.
The coding rows need the same care. On SWE-Marathon, the card says the GPT-6 Astra and Fable 5.1 scores came from Codex and Claude Code respectively, so "this row is not a controlled common-harness comparison", and Wonder Full was reported at xhigh effort rather than max. DeepSWE mixes Datacurve's runs for the Muse models with OpenAI's and Anthropic's own reports. None of this is hidden. It's just in a PDF most people won't open.
Our take: if you already pay Opus 5.5 rates, there's no index-level reason to switch to Wonder today, and Full isn't an option for production code paths at all. Where it might earn a trial is long, memory-heavy agent sessions, since Meta's internal MemoryBench is where it reports its clearest win over the Claude and GPT models it tested. That's Meta's own benchmark, though, and I'm not sure it'll hold up once outside groups test recall at 64K to 100K tokens. We'd wait for that before moving real traffic.
FAQ
Can I use Muse Wonder Full through the API?
No. Meta's model page lists muse-wonder-full-1.0 as available by subscription in Muse Code only. The Meta Model API serves muse-wonder-1.0.
How much does Muse Wonder cost?
$4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens on the Meta Model API, plus $2.50 per 1,000 web search queries. Meta doesn't bill the steering tokens it adds to each prompt.
Are Muse Wonder's weights available?
No. Both Wonder models are proprietary and can't be self-hosted. Meta's open-weight model in this family is the smaller Muse Glimmer, listed on our open-weight models tracker.
Does Meta train on what I send to Muse Wonder?
Not on the API: Wonder is offered in Standard mode only, and Meta says Standard content isn't used for training. Muse Code subscriptions are reported to default to Contributor mode, which allows training, so check that setting.
Sources: Meta, Muse Wonder announcement (12 October 2026), Muse Wonder model card, model page and Meta Model API pricing and rate limits; independent: Artificial Analysis, VentureBeat and TechCrunch. All read on 15 October 2026.
