Meta released Muse Wonder and Muse Wonder Full on October 12, putting its first frontier-class models into developers’ hands after months of positioning Muse as a consumer and workplace agent. The version most developers can call today is muse-wonder-1.0 through the Meta Model API; the more capable Wonder Full is limited to paid Muse Code plans.
That access split matters more than the headline benchmark chart. Muse Wonder brings a one-million-token context window, multimodal input, computer use and long-running agent features to Meta’s API, but it costs substantially more than Muse Spark. Wonder Full, meanwhile, has no public pay-as-you-go API price and cannot be selected by simply raising the reasoning setting on the standard model.
What Meta released
Muse Wonder is designed for complex software engineering and long-running agent work. According to Meta’s model documentation, it can take text, images, video, audio and PDF files as input, returns text, and supports reasoning levels from minimal through max. It uses the same Responses, Chat Completions and Messages endpoints as Muse Spark, so an existing integration can move between the two by changing the model ID.
Meta describes three capabilities as central to the release: Active Memory, which retrieves relevant details across conversations; Active Focus, which keeps an agent working through blockers; and stronger coding performance on sustained projects. The models can also call tools, delegate work to sub-agents and operate browser, desktop and mobile interfaces.
Muse Wonder Full is a separate higher-capability variant, not another reasoning tier. Meta’s documentation explicitly warns that setting Muse Wonder to max does not turn it into Wonder Full. Full is available only inside Muse Code on the $50-per-month Power Usage plan and the $200-per-month Maximum Usage plan. Meta has not listed token prices or API availability for it.
Muse Wonder pricing and the real cost of long context
The standard Muse Wonder API charges $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Unlike some long-context pricing systems, Meta says there is no higher rate after a prompt crosses a context threshold.
That does not make a million-token request cheap. Filling the context window with uncached input would cost about $4 before the model generates anything. A response containing 25,000 output tokens would add 50 cents. Reusing a stable cached prefix can cut the input portion sharply, but only the tokens served from cache receive the $0.20 rate.
Web grounding is billed separately at $2.50 per 1,000 searches, on top of model tokens. Teams building research or browsing agents should therefore track search calls alongside token use; a loop that repeatedly re-searches the same subject can create a bill that is not obvious from token counts alone.
Wonder is also Standard mode only on the pay-as-you-go API. In that mode, Meta says prompts, code, files, context and outputs are not used to train its models. The cheaper Contributor mode available for Muse Spark is not offered for Wonder through the API. Muse Code subscriptions do support Contributor mode and enable it by default, exchanging additional usage for permission to use submitted content and responses in future training. Developers handling private repositories or customer data should check that setting before a trial.
How Wonder compares with Muse Spark and Glimmer
Meta’s own guidance is straightforward: use Wonder when capability matters more than cost, including large multi-file changes, multi-hour tasks and sub-agent orchestration. Muse Spark 1.3 remains the better fit for high-volume classification, latency-sensitive tool loops and cheaper sub-agent work. Its standard API price is $1.25 per million input tokens and $4.25 per million output tokens, far below Wonder’s $4 and $20 rates.
Wonder is also a hosted model. It should not be confused with Muse Glimmer, Meta’s open-weight model for local deployment through runtimes such as vLLM, SGLang, llama.cpp and ExecuTorch. Meta has not announced open weights for either Wonder variant.
A sensible production architecture may use both hosted tiers: Wonder for planning, difficult repository changes and final review, then Spark for smaller delegated steps. Because the request formats are compatible, teams can test that routing without rewriting the application around a second API.
What the benchmarks do and do not establish
Meta presents Muse Wonder Full as competitive with leading models across agentic work, coding, computer use, memory and professional tasks. Its 13-page model card mixes independent results, provider-reported scores and Meta’s internal runs. Some comparisons use a common harness; others retain the agent product or tooling chosen by the benchmark vendor or competing model provider.
That makes the table useful for selecting workloads to test, but not a clean universal ranking. Meta acknowledges that internally generated competitor runs are best-effort and may not use prompts, tools or runtimes optimized for those models. Two prominent measures, the Personal Superintelligence Index and MemoryBench, are also internal evaluations.
The model card identifies further limits: multi-step reasoning can fail on unfamiliar scenarios, visual and spatial reasoning remains weak in some 3D tasks, performance may vary across languages, and results can shift with prompts, tools, quantization and the surrounding agent harness. For workflows involving legal, financial or health information, those caveats argue for task-specific evaluation and human review rather than relying on a composite score.
A practical evaluation plan
Developers considering Muse Wonder should begin with a small set of representative tasks rather than a broad chat comparison. A useful test should record success rate, wall-clock time, input and output tokens, cache hits, search calls, tool failures and the amount of human correction required.
- Repository work: give the model a multi-file change with executable tests and inspect whether it verifies its own work.
- Long context: repeat the same document task at increasing context sizes and watch retrieval accuracy, latency and cost.
- Agent recovery: introduce a failed tool call or changed requirement and measure whether the run recovers without losing prior constraints.
- Computer use: test in a disposable account or sandbox and require confirmation before purchases, messages, deletions or permission changes.
- Data handling: verify Standard or Contributor mode explicitly, then document which repositories, documents and customer records are permitted.
Meta’s model card recommends human confirmation for irreversible actions. That is especially important here: Wonder is marketed for agents that persist, operate software and act across personal or professional data. Better benchmark scores do not remove the need for narrow permissions, audit logs, spending limits and a clear stop mechanism.
Who should try it now
Muse Wonder is most interesting for teams already building coding agents, research systems or computer-use workflows where a cheaper model stalls on long tasks. Its compatible endpoints and clear model ID make a controlled test relatively easy. Teams running simple support, extraction or classification workloads will probably find Spark’s lower price more persuasive.
Wonder Full is harder to assess as a deployment option because it is confined to Muse Code subscriptions and lacks a public API rate. For now, it is best treated as a coding-product tier, not an API model that application developers can count on. The standard Wonder release is the part that can be evaluated, costed and integrated today.
Sources: Meta announcement, Muse Wonder product page, Meta Model API pricing, and the Muse Wonder model card.
