Meta Muse Wonder 1.0: what it does well, what it costs and who can use it

Picture a five-person web agency with a client's booking system to rebuild: new pages, a payment step, tests and a deadline next Friday. Until now, an AI coding assistant could help with each piece while a developer drove. Meta's new model is pitched at the next step: handing over the whole multi-hour job, letting the model split it among helpers, and getting a check-in when it needs a decision.

Meta released Muse Wonder and a more capable sibling, Muse Wonder Full, on October 12, 2026, calling them its most capable models and a step toward what it calls personal superintelligence. Muse Wonder is available to developers through the Meta Model API and Muse Code, Meta's coding agent for the terminal. Muse Wonder Full is only in Muse Code, on its two most expensive subscription plans. Meta says Muse Wonder will soon power its Muse personal agent app as well, starting with subscribers. ²³

Meta published a short introduction on its AI at Meta YouTube channel. It plays below in our player, so nothing loads from YouTube until you press play.

What Muse Wonder is built for

Meta describes Muse Wonder as built for complex software engineering and long-horizon agent work: jobs with many steps that, in Meta's words, can run for hours uninterrupted. It makes and checks changes across many files in real code repositories, plans the work, hands parts of it to sub-agents, calls tools in parallel and recovers from errors along the way. It can also operate browsers, desktop apps and mobile apps, which Meta calls computer use. ¹³

Two of Meta's terms are worth knowing. Active Memory means carrying important context about you across conversations and pulling up the details that matter for the task at hand. Active Focus means continuing toward a hard goal: trying other approaches, working through blockers, adapting to new information, asking for clarification when needed and checking in before important actions. That last behavior is the one to hold any agent to. ²

Muse Wonder takes text, images, video, audio and PDF files as input and returns text, with a context window of 1,048,576 tokens, about 1 million. A token is a small piece of text, roughly three quarters of a word. Developers can set its reasoning effort anywhere from minimal to max, trading speed and cost for more thinking. Meta lists the model as in preview. ³

Meta's own demonstrations lean toward everyday life as much as code: a playable 3D kitchen game built from images, an insurance claim, a Saturday running plan built by working across a browser and apps, personalized gifts and a wedding seating chart. They are short, chosen examples, so treat them as a description of intent rather than evidence of reliability. ²

How Muse Wonder scores against Claude and GPT

Meta's launch table compares Muse Wonder and Muse Wonder Full at maximum effort with Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 on 14 measures across agent work, coding and general ability. Several are run by outside groups: tests marked AA are run by Artificial Analysis, and Meta says Muse Wonder Full's leading DeepSWE and SWE-Marathon results come from independent third-party runs. The Personal Superintelligence Index is Meta's own internal composite of personal and knowledge work. ¹

Read across the whole table and the picture is closer than the launch language suggests. A Muse Wonder model has the top score in seven of the 14 rows, led by long coding work: on SWE-Marathon, a test of very long software projects, Muse Wonder Full scores 57.5 against 46.3 for Claude Opus 5.5. Claude Opus 5.5 has the top score in six rows, including ProgramBench, where it leads by more than 11 points, terminal coding and computer use. GPT-6 Astra leads one. Several of the leads, in both directions, are within a point. ¹

For a business, the rows that resemble real work matter more than the average. On AutomationBench-AA, which tests complete business workflows, Muse Wonder Full scores 45.4 against 41.2 for Opus 5.5 and 41.6 for GPT-6 Astra. Harvey's legal agent benchmark, built around source documents and detailed instructions, is low for every model but highest for Muse Wonder Full at 13.1. MemoryBench, a test of remembering context, is where the standard Muse Wonder edges ahead of everything, at 85.6. ¹²

An independent check: Artificial Analysis

Artificial Analysis runs its own tests on the models it tracks, and its Intelligence Index combines ten evaluations of professional work, coding, research and knowledge. It scores Muse Wonder at maximum effort 52.4 and Muse Wonder Full 55.8, against 57.6 for Claude Opus 5.5, 53.4 for Claude Fable 5.1 and 52.7 for GPT-6 Astra. By that measure Muse Wonder Full is near the top, and the standard model sits in a tight pack with GPT-6 Astra and Google's Gemini 4 Argon. ⁴

The chart shows where the money goes. Running the index tasks costs $4.52 with Muse Wonder, less than Claude Opus 5.5 at $5.98 but more than GPT-6 Astra at $3.26, for a similar score to Astra. GPT-6.1 Sol and Meta's own Muse Spark 1.3 score a few points lower for a fraction of the cost. ⁴

Speed is Muse Wonder's clearest independent advantage. Artificial Analysis measures a median of 145 output tokens per second on Meta's API, about twice GPT-6 Astra and Claude Fable 5.1 and half again as fast as Claude Opus 5.5. Meta's cheaper Muse Spark 1.3 is close behind at 139. These are rolling medians of live API measurements, so they move from day to day. Output speed is not the same as time to finish, because a model that reasons longer writes more tokens. But for an agent that works through hundreds of steps, faster steps add up. ⁴

What Muse Wonder costs

In the Meta Model API, Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Those are the same token prices Artificial Analysis lists for Claude Opus 5.5, and well under half of GPT-6 Astra's $10 and $50. Cached input is text the model has already seen in a recent request, such as a codebase or a brand guide an agent rereads on every step, which is why it is billed so much lower. ¹⁴

Muse Wonder Full is not sold by the token. It comes with a Muse Code subscription: the Power Usage plan at $50 a month, which Meta says gives 20 times the usage of its Everyday plan, or the Maximum Usage plan at $200 a month, with 80 times. Setting Muse Wonder's reasoning effort to max does not switch you to Full. ¹³

Who can use it, and the training setting to check

One setting matters for anyone handling client material. In Standard mode, Meta says your prompts, code, files and the model's responses are not used to train Meta models. Muse Code subscribers also get Contributor mode, which is switched on by default and gives more usage in exchange for letting Meta train future models on your content; you can turn it off in settings at any time. Through the pay-as-you-go API, Muse Wonder runs in Standard mode only. If your team uses a Muse Code subscription for client code, check that setting before the first session. ³

Meta also notes that model availability varies by country under its geographic use policy, so confirm the model is offered where your team works before you plan a project around it. ³

Muse Wonder or Muse Spark?

Meta's own guidance is simple, and it holds for choosing models in general. Use Muse Wonder for long, complex jobs where capability matters more than the cost per token: changes across many files, agent runs that last hours and coordinating sub-agents. Use the cheaper Muse Spark for high-volume, fast or cost-sensitive work, such as the sub-agent tasks themselves, sorting and labeling, and tight loops of tool calls. Both take the same requests on the same endpoints, so moving a job between them is a change of model name. ³

What to watch before you rely on it

Meta calls Muse Wonder its most aligned model yet, with its strongest safeguards so far against chemical, biological, radiological, nuclear and cyber misuse, but the full safety report has not been published. The model is in preview, Muse Wonder Full sits behind subscription plans and the Muse app has not switched over yet. Treat the coming weeks as a trial period rather than a migration. ²³

A fair trial uses a job you have already finished. Give Muse Wonder a closed task from the last month, such as a site change or a research brief, with the instructions your team had, read-only access where you can, and approval required for anything that sends or publishes. Record how long it ran, how often it asked for help, which errors it fixed on its own and what the run cost. Then run the same task on the model you use now. That comparison, not the launch table, tells you whether Muse Wonder belongs in your workflow.