Muse Wonder Is Meta's First Frontier Model
Meta's new flagship costs $4 per million input tokens, ships a long-term memory layer and powers the company's Muse agent.
In short
Meta released Muse Wonder, its first frontier model, on October 12, 2026, charging $4 per million input tokens and $20 per million output tokens through the Meta Model API.
At a glance
- Released October 12, 2026: Muse Wonder via the Meta Model API, Muse Wonder Full only for Muse Code subscribers.
- API pricing: $4 per million input tokens, $20 per million output tokens.
- TechCrunch reports benchmark parity with Fable 5.1 and GPT-6 Astra, and a lead over Fable, Opus and Astra on MemoryBench.
- Trained at Meta's Prometheus data center in central Ohio, which came online in 2026.
- Undisclosed: parameter count, context window and the license status of the weights.
Meta released Muse Wonder on October 12, 2026 — its first frontier model — priced at $4 per million input tokens and $20 per million output tokens on the Meta Model API. TechCrunch puts that in the same bracket as GPT-6 Sol or Opus 5.5. A company that built its reputation on giving weights away is now selling access by the token.
Parity at the top, a lead on memory
TechCrunch reports benchmark scores comparable to Fable 5.1 and GPT-6 Astra. On MemoryBench, Muse Wonder places ahead of Fable, Opus and Astra, and the report credits it with stronger results on complex work such as coding. No individual scores are published, so the margins cannot be checked.
Active Memory is the pitch
The headline capability is Active Memory: the model is built to retain and synthesize what a user feeds it, rather than resetting at the end of a session. Meta is aiming this at personal assistants that stay on a task for days, and the same model powers the Muse agent.
Research scientist Kavya Kopparapu says the team worked hard on letting the model remember "what's important" to a user across days, weeks and months.
Who gets Muse Wonder Full
Only the smaller Muse Wonder is served through the API. Muse Wonder Full is restricted to Muse Code subscribers and is not exposed as an endpoint. Meta has not said what the subscription costs or where Full's extra capability actually lies.
Built in Ohio
Training ran at Meta's Prometheus data center in central Ohio, a site that came online in 2026. This is the first Meta model generation to come out of that build.
What stays unverified
Parameter count, context window and license terms are not public. The preview text of the Wired story points to new features meant to keep agents from going off the rails, but we could not load that article, nor the Golem report behind its consent wall. We are therefore not describing those guardrails as confirmed, and will add them once a source we can read states them.
FAQ
How much does Muse Wonder cost per million tokens?
$4 per million input tokens and $20 per million output tokens on the Meta Model API, which TechCrunch describes as comparable to GPT-6 Sol or Opus 5.5.
What does Active Memory do in Muse Wonder?
It lets the model keep and combine information a user supplies across long-running work instead of starting fresh each session, which is what Meta points to for assistants that run for days.
Is Muse Wonder open weights?
Unconfirmed. The one report we were able to read states no license, parameter count or context window, so treat the weights question as open rather than answered.
