Meta has spent two years telling the world it would catch up with OpenAI, Anthropic and Google. On October 12 it shipped the model that is supposed to prove it. Muse Wonder is live, priced, and already serving traffic. What it does not have yet is the thing that usually comes with a frontier launch: a pile of independent test results and a research post explaining how it works.
The short version
- Meta Superintelligence Labs released Muse Wonder and a larger Muse Wonder Full on October 12
- The standard model costs $4 per million input tokens and $20 per million output tokens, with a context window of about 1.05 million tokens
- Weights are closed, a notable shift for the company behind Llama
- Early independent scoring puts it level with the leaders but not ahead, and most headline claims are Meta’s own
What Meta actually released
There are two versions. Muse Wonder is the one developers can call through Meta’s model API, and it also shows up on OpenRouter. Muse Wonder Full is the bigger sibling, and it is not sold by the token at all. You reach it through Muse Code, Meta’s terminal coding agent, on a paid monthly plan.
Meta describes the model as built for long, messy agent work: it plans, hands pieces to sub-agents, calls tools in parallel and recovers from errors. It takes text, images, video, audio and PDFs, and it can operate browsers and desktop or mobile apps. Two features carry the marketing. Active Memory is meant to carry context across conversations over days and months. Active Focus keeps the model pushing toward a goal instead of stopping after the first failed attempt.
| Spec | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Access | Meta API and OpenRouter | Muse Code subscription only |
| Input price | $4.00 per million tokens | No per-token price |
| Output price | $20.00 per million tokens | No per-token price |
| Cache read | $0.20 per million tokens | Not listed |
| Context window | About 1.05 million tokens | Not confirmed |
| Weights | Closed | Closed |
Reports disagree on what the subscription costs. Some list $50 and $200 tiers, while one outlet says up to $100 a month. Check Meta’s pricing page before you budget around it.
How it scores so far
The most useful number so far comes from Artificial Analysis, which runs its own Intelligence Index rather than relying on vendor charts. On that scale Muse Wonder Full lands at 56 and the standard Muse Wonder at 52. That ties Claude Sonnet 5.5, edges past GPT-6 Astra, and sits two points behind Claude Opus 5.5.
Meta’s own charts tell a friendlier story. It reports 79.0 on DeepSWE 1.1, a coding benchmark, against roughly 74 for both Opus 5.5 and GPT-6 Astra. It also leads on MemoryBench, but Meta built that test itself, which is exactly the kind of result to hold at arm’s length. To its credit, Meta’s own tables still show Opus 5.5 ahead on several rows, including OSWorld and Terminal-Bench.
The strange part: no research post
Every earlier Muse model came with an announcement on Meta’s research blog. This time the blog’s latest entry is still Muse Spark 1.1 from July 9. The news traveled through a Zuckerberg post on Threads instead, while the model quietly started serving real traffic on OpenRouter. Meta did not answer questions about the gap. Early testers have also noted that hands-on demos look less impressive than the benchmark slides.
None of this means the model is weak. It does mean the usual evidence is missing: no technical report, no safety write-up, and no third-party evaluations beyond the first index scores.
Why the closed weights matter
Meta built its AI reputation on open releases. Muse Wonder breaks that pattern, which fits the direction we saw when Meta pitched an agent that can email, shop and pay for you. A company that wants to sell a premium agent subscription has less reason to hand the engine to everyone for free.
What is still unverified
- Benchmark gains. Most come from Meta’s own slides, and different outlets report different rankings depending on the test
- Subscription pricing. Coverage conflicts on the exact monthly tiers
- Real agent reliability. Long-running browser and desktop tasks are hard to judge from launch demos
How it fits the wider race
Meta now has a model in the same band as the leaders, but the leaders are not standing still. Chinese labs are also closing in, as we covered when DeepSeek came within 3% of the best US model. A tie at the top is a respectable first result, though not the leap Meta spent billions chasing.
What to watch next
- Independent testing. More third-party benchmarks over the next two weeks will show whether the launch slides hold up
- A technical report. Meta still owes one, and its absence will keep fueling doubts
- Developer uptake. At $4 and $20, Muse Wonder is priced like a premium model, so it needs to win on agent tasks to earn a place
For now, Muse Wonder is best read as a credible entry rather than a victory lap. Meta is in the room with the leaders. Whether it can stay there will depend on what independent testers find once the launch buzz fades.
Sources and further reading
- Forkast: Meta Ships Muse Wonder, Its First Frontier Model, and Leaves the Research Blog Empty
- AI Weekly: Meta releases Muse Wonder, its first frontier-class AI model
- Startup Fortune: Zuckerberg says Meta’s Muse Wonder AI has reached the frontier and keeps building anyway
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
