Meta released Muse Wonder and Muse Wonder Full on October 12, 2026. Wonder is available through the Meta Model API; Full is restricted to eligible Muse Code subscriptions. Selecting maximum reasoning on API Wonder does not select Full. For developers comparing agent performance, that access distinction determines which result they can reproduce.
The launch announcement positions the pair for sustained coding and professional workflows, with Active Memory and Active Focus. Those are vendor capability descriptions. An application still needs its own record of completed tasks, failed actions and recovery behavior before assigning unattended work.
The model and the reasoning setting are separate choices
Meta’s model availability documentation identifies muse-wonder-1.0 as the API model. Its context window is 1,048,576 tokens, with text, image, video, audio and PDF input and text output. Wonder Full is available in Muse Code’s Power Usage and Maximum Usage plans, rather than as a pay-as-you-go API option.
The Muse Code subscription page currently lists those plans at $50 and $200 per month. A subscription price cannot be converted directly into a Full token rate: the product allocates usage, and a difficult task may consume more of that allowance than a short request. Verify the eligible model in the actual session before comparing outputs.
This also creates an evaluation boundary. A Full benchmark result does not establish the behavior of a workflow running API Wonder, even if both use the highest reasoning setting. Record model identity, effort, available tools and the agent application separately. A coding product can supply orchestration and context management that a bare API call does not provide.
Hidden reasoning belongs in the output bill
The reasoning guide says internal reasoning tokens count toward both the output budget and billed output tokens. They do not appear as ordinary response text. Higher effort can therefore raise cost and time even when the final answer remains short. The none setting is unsupported for Wonder and returns an error; the documented range begins at minimal.
The current API pricing is $4 per million uncached input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Consider a hypothetical request with 100,000 uncached input tokens and 10,000 total billed output tokens, including reasoning. Its token charge is $0.40 + $0.20 = $0.60. If 80,000 input tokens qualify for caching, the same accounting becomes $0.08 + $0.016 + $0.20 = $0.296.
Those are calculations, not measurements of Wonder’s consumption. They exclude tools and other charges. More reasoning can overwhelm input savings: an additional 20,000 billed output tokens adds $0.40. For budgeting, collect the API’s usage fields rather than counting words visible on screen.
An agent’s useful denominator is accepted work. If ten attempts cost $6 in total and eight produce accepted changes, the observed token cost is $0.75 per accepted change. Include abandoned runs and retries in that numerator. Otherwise the longest failures disappear from the reported economics.
Benchmark provenance matters more than a single ranking
Meta’s model card reports max-effort evaluations. Its methodology distinguishes controlled internal coding comparisons from external results that retain the reporting vendor’s application or setup. It also describes selecting among internally validated, provider-reported and official leaderboard scores when comparable third-party results are unavailable. These are different evidence paths, not one uniform experiment.
A reproducible deployment comparison should fix the repository revision, permitted tools, success criteria and retry budget. Report elapsed time, token use and accepted results for each model. Keep product-level Full trials separate from API Wonder trials, and mark changes in the surrounding agent software. A one-million-token capacity figure alone does not measure dependable recall across that window.
Check the data mode before connecting a repository
The models documentation says Standard-mode content is not used to train Meta models. Muse Code subscribers also have Contributor mode, enabled by default, which exchanges permission to use prompts, code, files, context and responses for more usage. It can be disabled in settings. Wonder’s pay-as-you-go API path does not offer that mode.
For a team, model selection and data-mode selection belong in the same deployment record. Full access is useful only if its subscription workflow, usage allocation and content treatment fit the intended repository. The API route offers a clearer per-token experiment, but it evaluates Wonder, not Full.
Prepared with AI assistance from the cited primary documentation. Cover: AI-generated conceptual illustration. EyesTech did not benchmark these models for this article. Prices and availability checked October 13, 2026.
