Latest in AI
Muse Wonder: API Pricing, Context Window and Full Model Access
Meta's new agent models bring a larger ambition to Muse, but the API, subscription and privacy choices need a closer look.
Meta introduced Muse Wonder and Muse Wonder Full on October 12, with the standard model available in preview through its API and Muse Code. The release targets coding and long-running agent work, combining memory across conversations with the ability to keep working through difficult tasks.
The useful question is which version you can actually use. Wonder Full has a different access path, and the subscription route comes with a data-use setting worth checking before pointing an agent at a private repository.
Muse Wonder API pricing
Meta's pricing documentation lists these pay-as-you-go rates in US dollars per million tokens:
| Usage | Muse Wonder | Muse Spark 1.3, Standard mode |
|---|---|---|
| Input | $4.00 | $1.25 |
| Cached input | $0.20 | $0.15 |
| Output | $20.00 | $4.25 |
There is no additional long-context premium. Web search grounding costs $2.50 per 1,000 queries on top of token charges. Wonder does not have a discounted Contributor-mode API option.
For a simple illustration, one million uncached input tokens and 100,000 output tokens would cost $6 with Wonder, before tools. The same token counts would cost $1.675 with Spark 1.3 in Standard mode. That comparison holds token usage constant; it doesn't tell us what either model will spend to finish an actual task.
A stronger agent can sometimes justify a higher token price by needing fewer retries or less human repair. It can also spend more time exploring alternatives. For a team evaluating the switch, cost per accepted change is a more useful measurement than the price of one answer.
Context window and API availability
The model documentation labels Muse Wonder a preview and specifies a 1,048,576-token context window. It accepts text, images, video, audio and PDFs, and returns text. The API model identifier is muse-wonder-1.0.
Wonder is available through pay-as-you-go access and every Muse Code subscription plan. Wonder Full is limited to Muse Code's Power Usage and Maximum Usage subscriptions; it is not available through the public Model API. Setting the standard model's reasoning effort to max does not turn it into Full.
That distinction matters when reading benchmark comparisons. A result for Full should not become an expectation for an application calling the standard API model.
What Muse Wonder Full costs
The Muse Code product page lists Power Usage at $50 a month and Maximum Usage at $200 a month. Those are the two plans Meta identifies for Full access. The lower Everyday Usage and High Usage tiers are $5 and $15 a month respectively.
Check Contributor mode before using a subscription for confidential work. Meta's documentation says this setting is on by default for Muse Code subscribers. It provides more usage in exchange for permission to use submitted content and responses to train future models. Users can turn it off; Standard mode does not use that content for model training.
This is a meaningful purchasing decision for a company. A developer choosing a plan is also choosing how code, prompts and attached context may be used. The usage allowance and the data setting need to be reviewed together.
What Active Memory and Active Focus mean
In its research announcement, Meta describes Active Memory as retaining important context across conversations and retrieving it when relevant. Active Focus describes sustained work: trying other approaches, adapting to new information and asking for clarification or approval when needed.
Consider a coding agent working on the same product across several sessions. Remembering a deployment constraint could prevent it from repeatedly proposing an unsuitable dependency. Staying focused could help it work through a failed test instead of stopping at the first plausible patch. These are the kinds of behaviours the release is trying to improve, rather than a promise that every long task will finish correctly.
Meta reports strong results on agent, memory and coding evaluations, including MemoryBench and SWE-Marathon. Those comparisons need to be read with their test setup and model variant attached. They do not establish reliability on a team's own repository or permission boundaries.
What still needs testing
The consumer Muse rollout is still ahead: Meta says Wonder will reach the personal-agent product first for subscribers. The developer release should not be read as confirmation that every Muse user already has the new model.
The research announcement also says an additional safety report will follow. For now, the practical evaluation should include whether an agent recalls the right details, abandons stale instructions and stops before consequential actions. Remembering more is useful only when the system can distinguish what still applies.
I think that is the most important test of this launch. If a team can give an agent a difficult job, return later and find a verified result with understandable decisions, the higher inference bill may be reasonable. If the team still has to reconstruct every step, a larger model has only moved the supervision work around.
I'm Paras. I work at the intersection of AI, engineering and distribution.
If you're building an AI product and want help turning technical updates into launch stories and founder content people understand and share, I'd be happy to work with you.