Meta's best Muse model is not the one you can buy via API

Meta is back at the frontier. Its best model is not the one you can put in your own product.

On October 12, Meta launched two models from Meta Superintelligence Labs, the AI research organization led by former Scale AI CEO Alexandr Wang: Muse Wonder and Muse Wonder Full. Both are built for agent work: long tasks across multiple apps, keeping context, recovering from failure. Meta's framing centers on two capabilities, Active Memory (carrying relevant context across conversations) and Active Focus (pushing toward a goal, revising plans, asking permission before consequential actions).

The split is what matters. Wonder is available through the Meta Model API under the model ID muse-wonder-1.0, and through Muse Code. Wonder Full is available only inside Muse Code, on the Power Usage plan at $50 a month and the Maximum Usage plan at $200 a month. There is no pay-as-you-go API price for Wonder Full. If you want to embed Meta's strongest model in your own software, you currently cannot.

Neither model ships with publicly downloadable weights. That is a clean break from the Llama era, when Meta's downloadable models let organizations run them on their own infrastructure. Wonder is a hosted service. Meta's API terms give customers ownership of submitted inputs and generated outputs, but access runs through Meta's service, not an open-weight license.

The pricing for the model you can buy: $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million. Web search through the API costs $2.50 per 1,000 queries. Meta charges the same token rate regardless of how much of the roughly one-million-token context window a request uses. The rate matches Claude Opus 5.5 exactly, and undercuts GPT-6 Astra's $10/$50. Wonder is in preview.

The benchmark story needs the same split applied. Independent benchmarking organization Artificial Analysis scored Wonder Full at 56 on its Intelligence Index, tying Claude Sonnet 5.5, ahead of GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 at 53 each, and behind Claude Opus 5.5 at 58. The standard Wonder scored 52, one point behind GPT-6 Astra. Most of the headline numbers belong to the model you cannot call from the API.

Meta's own model card adds detail, with the usual caveat that its table mixes Meta-run evaluations, vendor-supplied scores, and third-party results, and that some coding comparisons use different tools. On SWE-Marathon, a long-running coding benchmark, Wonder Full scored 57.5 against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra. On the Harvey LAB-AA legal benchmark it scored 13.1 against 4.2 for Opus, though the low absolute scores show how far agents remain from reliably producing satisfactory legal work. Opus still leads on Terminal-Bench and on health questions. Neither set of numbers replaces testing on your own tasks. Artificial Analysis' cost-per-task estimate makes the same point from the other side: $4.52 per average Intelligence Index task for Wonder, versus $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra. Per-token price is not cost per finished task.

One more split, and it is the one to check before anyone on your team subscribes. Pay-as-you-go Wonder API access runs in Standard mode: Meta says submitted content is not used to train its models. Muse Code subscriptions default to Contributor mode, which permits Meta to use submitted prompts, code, files, context, and outputs to improve future models. You can turn it off, but the change applies only to future submissions. Meta's terms prohibit submitting confidential or sensitive content. If your team pastes proprietary code into Muse Code on the default setting, you have a data-governance problem, not a pricing problem.

The consumer rollout comes later: Wonder will power features in Meta's Muse personal agent, with subscribers first. Meta gave no date.

So the decision is narrow. Test Wonder on 30 to 50 of your own real tasks next to your current model. Compare success rate and cost per finished task. Switch only the workloads where it clearly wins, and keep a fallback ready. Do not buy a $50 Muse Code plan expecting API access to Wonder Full. Do not read Wonder Full's benchmark scores as Wonder's. And check the Contributor mode default before the first paste.

Related: Meta's agentic commerce split covered the other half of the Muse strategy, and the Opus 5.5 price cut set the rate Wonder now matches. The GPT-6.1 Sol pricing shows where OpenAI is pushing on speed instead.