On Monday October 12, 2026, Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full, two frontier-tier models trained at the Prometheus data centre in central Ohio. Muse Wonder is available on the Meta Model API at $4 per million input tokens and $20 per million output tokens, matching GPT-6 Sol and Claude Opus 5.5. Muse Wonder Full — the larger and more capable of the two — is not on the API at all. It is available only to subscribers of Muse Code, Meta's paid coding product, and will reach the Muse consumer app for paid subscribers "soon" (TechCrunch, The Decoder).

This is the first time Meta has shipped a model that lands in the top tier of independent third-party benchmarks. It is also the first time Meta has priced a model at frontier parity rather than at the Chinese-open-weight floor. The company that spent 2025 and 2026 persuading the market it was serious — hiring Alexandr Wang out of Scale AI for a reported fourteen billion dollars in August 2025, pouring compute into the Prometheus and Hyperion sites, calling Muse Spark an "appetiser" at Bloomberg Tech in June — has shipped the main course. The main course costs twenty dollars a million output tokens and the better plate is at a different table.

The two-model split is a product decision, not a technical one

Meta's blog post, quoted by TechCrunch, calls Wonder and Wonder Full "the next step on our scaling ladder and a major leap toward personal superintelligence." The two models share the same release day, the same training site, the same architecture family, the same 1-million-token context window, and the same multimodal input stack across text, image, video and audio. What splits them is distribution, and the split is sharp.

Muse Wonder is on the API. A developer with a Meta API key can call it tomorrow. The price is $4 / $20 per million tokens, which is the same tier as Opus 5.5 ($4 / $20) and GPT-6 Sol ($4 / $20) and above GPT-6.1 Sol ($0.72 per task on the Artificial Analysis Intelligence Index) and Gemini 4 Argon ($1.99 per task). The Decoder's accounting puts Muse Wonder's per-task cost at $4.52, slightly above Astra and slightly below Opus 5.5.

Muse Wonder Full is not on the API. The only way to use it today is inside Muse Code, Meta's coding subscription, which starts at a monthly rate the company has not yet published. The consumer Muse app will add it for paid subscribers at some unspecified date. There is no "Full" pricing on the Meta Model API because there is no "Full" on the Meta Model API.

That is Anthropic's Fable 5 play, run again by a different company. Keep the best model inside the product. Let the API have the second-best. Capture the subscription revenue directly instead of selling the margin to whoever is building a wrapper on top. The one thing Meta added to the pattern is that the subscription in question is a coding product — not a general consumer chat subscription. Which means the single most powerful Meta model as of October 12 is accessible only to people who are already paying Meta to help them write software.

The benchmarks say frontier-adjacent, not frontier-leading

On the independent third-party scoreboards, the two models land in two different places.

The Artificial Analysis Intelligence Index v4.3 composite puts Wonder Full at 56, Wonder at 52, and the previous Muse Spark 1.3 at 48. For context, that puts Wonder Full in the band with Opus 5.5 and GPT-6 Astra. On the Vals AI Index, which weights real-world task performance, Wonder Full (Max setting) ranks fifth overall. It beats every OpenAI model on the index. Only Anthropic models — Opus 5.5 and Fable 5.1 — are above it. Wonder is tenth.

Coding is where Meta's six-month compute binge shows up most clearly. Terminal-Bench scores climbed from 33 percent to 57 percent on agentic coding tasks between the Spark 1.3 predecessor and the Wonder Full. On the Coding Agents Index, Wonder Full posts 68, tying Claude Code running on Sonnet 5.5. Wonder is 63 and the predecessor was 54. For a lab that was shipping open-weight Llama derivatives twelve months ago, this is a step change. For buyers evaluating what to switch to, it is now a real option on the enterprise-coding beat where Anthropic has been running away with share for most of 2026.

Mathematical reasoning posts the single most striking number in the release. On ProofBench, which measures theorem-style problem solving against a 100-problem set, Wonder Full solves 94. The predecessor solved 58. The top models in the field solve 99 or 100. A jump of thirty-six problems in a single model generation is on the order of what OpenAI claimed on the 719-manuscript release last week, and Meta did it with a model that is also available for ordinary agent work.

The soft spots are legible. The AA-Omniscience factual-knowledge benchmark gives Wonder Full 30 and Wonder 25, against Opus 5.5 at 46, GPT-6 Astra at 43, and Gemini 4 Argon at 42. That is a knowledge gap, not a reasoning one, and it is the kind of gap that a scaling-limited lab closes by throwing more corpus and more compute at a problem a competitor has already solved. The hallucination rate on Wonder Full is 23 percent, the second-lowest in the band of models that score 45 or more on omniscience — Gemini 4 Argon still leads at 15 percent. Finance and law benchmarks regressed by up to two points relative to Spark 1.3, which is the kind of result that gets a specific team at Meta a specific meeting.

The honest reading: Meta joined the frontier tier. It did not take the top of it.

Active Memory is the product bet

The feature Meta picked to lead the launch is Active Memory. In the words of research scientist Kavya Kopparapu in Meta's launch video, quoted by TechCrunch: "We put a tremendous amount of effort into making sure your model can remember what's important to you." The second named feature is Active Focus — the model's ability to persist on hard tasks, try alternative approaches, and ask clarifying questions.

Active Memory is the industry term for remember what you said five minutes ago. OpenAI shipped it as a ChatGPT product feature in 2024. Anthropic shipped project-level memory in Claude in 2025. Google rolled Gemini memory into Workspace during the summer. Meta's bet is that persistent memory stops being a chat convenience when the agent on the other side is a coding assistant that holds context across days of a project, or a consumer agent that holds context across the appointments and messages in your life. On MemoryBench, Meta's own benchmark, Wonder beats Fable 5.1, Opus 5.5 and GPT-6 Astra, per Meta. That is a Meta-authored result on a Meta-sympathetic benchmark and it will travel exactly as far as independent re-tests take it.

The framing matters more than the first-round number. If memory is the differentiator Meta leans on across the next year of releases, this launch is the opening move in making the API-vs-subscription split structural — the Full-tier model always gets the long-horizon memory capability because the long-horizon memory capability requires a running state that is easier to maintain inside a subscription product than inside a stateless API call. The company that gets to own your context, the pitch goes, gets to own the agent too.

What to watch

1. Independent reproduction of the memory and coding benchmarks. Artificial Analysis and Vals AI will do their re-tests in the next ten days. The AA-Omniscience 30 and the ProofBench 94 are the two numbers whose confirmation or climb-down shapes the one-month verdict on Wonder Full.
2. Whether Muse Wonder Full ever reaches the API. Meta has not said it will. Anthropic shipped Fable 5 the same way and Fable 5 lives on inside the Max subscription a quarter later. If the pattern holds through Q1 2027, the API is no longer where the best models live and the enterprise-procurement conversation has to be re-written around subscription seats rather than token meters.
3. Muse Code pricing and seat economics. Meta has not published a Muse Code price. The subscription-only distribution of Wonder Full only resolves into a buyable product once that number lands. If Muse Code is priced below Claude Code's $200/month top tier, Meta's play is a direct raid on Anthropic's coding seat. If it is priced above, Meta is betting that Active Memory is the thing enterprise coding teams were missing.
4. The next Meta release's open-weight status. Muse Spark shipped closed. Wonder and Wonder Full have no open-weight announcement. Alexandr Wang has repeatedly said Meta intends to return to open weights at some point. The next release's licence will be the live reading of whether that promise survived contact with the leaderboard.

Meta spent fourteen billion dollars on an engineering reorganisation and a data centre named after the Greek titan who stole fire from the gods and gave it to humanity. The first thing it did with the fire was price it at Opus-5.5 rates and keep the better torch inside a coding subscription. That is a reasonable commercial read of where the market is in October 2026. It is also an admission that the open-weight pitch that defined Meta's AI identity through Llama is now a line in a press release and not a fact about the product.