Muse Wonder: Meta's First Frontier Model, and the Access Quirk Nobody Should Miss

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, 2026 — the company's first models it calls frontier-class. Built for personal and professional AI agents, they are Meta's answer to Claude Opus 5.5 and GPT-6 Astra. But there is one detail every developer should know before writing integration code: the better model is not on the API.

Two Models, Two Doors

• Muse Wonder (muse-wonder-1.0) is on the Meta Model API and inside Muse Code, Meta's subscription coding and agent product.
• Muse Wonder Full is the larger, more capable variant — and it is not sold through the API at all. It is reserved for Muse Code subscribers on the Power Usage ($50/month) or Maximum Usage ($200/month) plans.

Setting effort to max on the API's Wonder does not give you Wonder Full. Meta treats them as separate models. If your workload needs the strongest version and you are on pay-as-you-go, the door is closed — you need the subscription.

The Spec Sheet

Both models accept text, images, video, and audio as input and generate text output, with a shared context window of roughly one million tokens (1,048,576 on the API). The toolkit includes function calling, JSON-schema outputs, built-in web search, and six effort settings from minimal to max — so you can trade cost against capability per call.

Meta trained both at its Prometheus data center in Ohio, with a larger base model and extra reinforcement learning over the earlier Muse Spark family. (Muse Spark 1.3 scored 48 on the Artificial Analysis Intelligence Index six weeks prior; Wonder jumped to 52.)

Active Memory and Active Focus

The pitch centers on two named features aimed at agent builders:

• Active Memory carries what matters across conversations — key facts, preferences, project context — without re-explaining everything each session. Meta claims it leads its own internal MemoryBench.
• Active Focus is built for long-running tasks: sustaining effort, retrying different approaches when stuck, and checking in before doing anything important.

If you build agents that run for minutes or hours, these are the features to benchmark first — Meta's memory claims are vendor-reported and have no independent verification yet.

Pricing and Benchmarks

API pricing: $4 per million input tokens, $0.20 cached input, $20 per million output — the exact headline rate as Claude Opus 5.5. Meta is deliberately matching Anthropic's flagship price, not undercutting it.

Independent numbers: Artificial Analysis scores Muse Wonder at 52 and Wonder Full at 56 on its Intelligence Index, versus 58 for Claude Opus 5.5. Meta's best lands second in the world, two points behind the leader and ahead of GPT-6 Astra (53). Genuinely impressive for a first attempt — and still the trailer, not the movie.

What to Watch

Wonder is in preview at launch, so expect rough edges, possible regional availability limits, and changing behavior. The full safety report is "coming soon" per Meta. And there is a strategic footnote worth reading: after two years of insisting on open models, Meta's frontier play is closed and paywalled — the Llama-era pricing is gone.

If you build long-running agents or coding workflows and want a 1M-token model at flagship rates, Muse Wonder is worth a serious look — one independent cost test even found it cheaper per completed task than Anthropic's best, though that is a single data point, not a trend. Wait if you need a stable, non-preview model or the strongest variant without a $50/month commitment.

This is a condensed explainer. The full breakdown — the complete Wonder vs Wonder Full comparison, the pricing details, pros and cons, who each tier is for, and an FAQ — is on NeurAITool: Muse Wonder Explained: Meta's First Frontier AI Model. Tracking the rest of the leaderboard? See our ChatGPT vs Claude vs Gemini comparison.