Meta ships Muse Wonder frontier models for agents at $4 per million input tokens

Meta shipped Muse Wonder and Muse Wonder Full on October 12, 2026. These are its first frontier models, built for long-running personal and professional agents. Muse Wonder is live on the Meta Model API and in Muse Code today. Wonder Full sits only on Muse Code’s $50 and $200 plans.

The API price for Muse Wonder is $4 per million input tokens, $0.20 for cached input, and $20 per million output. No long-context premium. Context window is 1,048,576 tokens. Inputs include text, images, video, audio, and PDF. Output is text. Rate limits run 3,000 requests and 4 million tokens per minute per team. The model is in preview. You can drop the model string muse-wonder-1.0 into any OpenAI-compatible client pointed at https://api.meta.ai/v1.

Wonder Full has no public per-token price. Meta limits it to Muse Code subscribers on the Power Usage plan at $50 per month and the Maximum Usage plan at $200 per month. Those plans give 20x and 80x the Everyday Usage allowance, respectively. Everyday starts at $5 per month for 10–50 prompts every five hours. High Usage sits at $15 for 5x that. Contributor mode is available on subscriptions and trades usage for training data rights. It is off by default on pure API pay-as-you-go for Wonder. In Standard mode your prompts and outputs are not used for training.

Meta trained both models at its Prometheus data center. It calls out two training properties. Active Memory carries relevant context across conversations and pulls details back when a later task needs them. Active Focus keeps the model working through blockers, trying alternate approaches, asking for clarification, and checking before important actions. Meta tested these on MemoryBench and FocusBench. On MemoryBench the models score higher than Claude Fable 5.1, Opus 5.5, or GPT-6 Astra in Meta’s runs. The properties are baked into the training, not toggles you flip at inference.

The models take multimodal prompts and support computer use, tool calling, structured output, and search grounding. Search grounding costs an extra $2.50 per 1,000 queries. Meta positions them for multi-file coding changes, long-horizon agents that plan and recover, and everyday computer-use tasks. Examples in the launch materials include building a playable 3D kitchen game from image inputs, planning a Saturday run with map data, generating custom swag from Instagram photos, and seating a wedding under complex constraints. The research post shows agents working directly in a browser and across apps.

On Artificial Analysis Intelligence Index v4.3 at max effort, Muse Wonder Full scores 56 and Muse Wonder scores 52. That places Wonder Full two points behind Claude Opus 5.5 and ahead of several GPT-6 variants in the reported tables. Meta’s own model card lists stronger numbers on SWE-Marathon (57.5 for Wonder Full at xhigh), DeepSWE (79.0), and Harvey LAB legal work. Independent runs cited by third parties show the same direction on long coding tasks, with Opus still leading some terminal and computer-use rows. A full safety report is still forthcoming. Meta says the models refuse harmful acts, avoid leaking private information, and keep defensive cyber capabilities available under its Superintelligence Scaling Framework.

Availability is straightforward. Developers can point an OpenAI-compatible client at the Meta Model API with model muse-wonder-1.0 and a MODEL_API_KEY. Muse Code users on the lower plans get Wonder; Power and Maximum get Wonder Full. Meta says Wonder will power more of the Muse consumer agent soon, starting with subscribers. The Muse app itself launched in September on Muse Spark. No open weights. This is closed, like current Claude and GPT frontier models. Setting reasoning effort to max on the API does not unlock Wonder Full.

For an indie founder running agents on customer support, sales follow-up, or internal ops, the practical difference is the long-task behavior and the price relative to Opus-class models. At $4/$20 the input is cheaper than several Opus and Astra listings that sit at $10/$50. Cached prefixes drop further to $0.20. You still pay for every tool call and for search. On a $50 Muse Code plan you can run Wonder Full without metering every token, but the five-hour windows and weekly caps still exist. Contributor mode lowers the effective cost if you accept the training clause. A cold turn with 60k input and 3k output lands around $0.30 on Standard Wonder before cache hits.

The launch sits fourteen months after Meta’s internal AI overhaul and six weeks after Muse Spark 1.3. Shares moved about 4 percent on the day. Meta is selling the models as the next step on its scaling ladder toward personal superintelligence, with Active Memory and Active Focus as the concrete mechanisms for keeping agents on multi-day work. The API is open for self-serve testing today. Wonder Full requires the subscription. Safety details land later. Those are the numbers and the gates.

If you already have a Meta Model API key, the model string is muse-wonder-1.0. If you are on Muse Code and want the higher variant, check whether your plan is Power Usage or Maximum Usage. Both variants share the 1M context and the same multimodal input types. The split is capability and access, not a different interface. Meta’s docs note that setting reasoning effort to max on the API model does not unlock Wonder Full. You have to be on the right subscription. Rate limits stay the same across the Wonder series at 3,000 requests and 4M tokens per minute.

Muse Code itself is a terminal agent that Meta introduced earlier. Installing the CLI is free. You pay either pay-as-you-go or through the monthly plans. Subscriptions meter requests inside five-hour windows rather than raw tokens for the included usage. Hitting the limit means upgrading or waiting for the refresh. Extra API keys on the same account bill separately. The Power and Maximum plans also unlock early access and higher file upload limits. For a solo founder who already runs Muse Code on customer codebases, Wonder Full is the incremental spend if the long-horizon coding scores hold up in your own repos.

Compared with Muse Spark 1.3, which still sits at $1.25 input and $4.25 output on the same API, Wonder is more than three times the input price. The jump is intentional. Spark was the lighter model for lower latency. Wonder is the one Meta is pointing at multi-hour agent jobs and large-repo changes. If your workload is short chat or simple retrieval, Spark remains cheaper. If you need the model to stay on a coding or ops task for an afternoon, the new pricing is the rate card Meta chose to match Opus-class models.

That covers the ship. Prices and limits come from Meta’s pricing page and model docs as of October 13. Benchmarks mix Meta’s card with third-party index scores. The product is live. You can test the API today without a subscription. Wonder Full waits for the $50 or $200 plan.