Meta launches Muse Wonder, with Full available through Muse Code

Meta released Muse Wonder and Muse Wonder Full on 12 October. Wonder is available through the Meta Model API and Muse Code; Full is available to Muse Code subscribers. A rollout into the consumer Muse service is still described as forthcoming, starting with subscribers. Those are three different access arrangements, rather than one model becoming available everywhere at once.[1]

The access route determines the first practical decision. An API builder can try Wonder in an existing application; a developer seeking Full needs a Muse Code subscription. Consumer availability will follow its own rollout, so the API launch alone does not establish that a particular chat account has been upgraded.[1]

The independent improvement is visible in the results

Artificial Analysis tested the models at maximum reasoning effort. Its launch assessment reports gains on both a broad intelligence index and a terminal-based coding benchmark. The latter is particularly relevant to a coding agent: it tests work in a computer environment rather than merely asking a model to explain code.[4]

That is evidence of a stronger offering, not a requirement to buy the highest-scoring version for every job. A short classification task, an extended software repair and a question about a personal document place different demands on an application. The useful commercial choice depends on which of those jobs dominates.

The model card lists text, image, video and audio inputs, text output, and a context window of one million tokens. A large input allowance can accommodate more material in a request; it does not mean an application automatically remembers everything from previous conversations.[3]

Memory and sustained work are separate capabilities

Meta describes Active Memory as bringing relevant information forward across conversations. Active Focus is its term for maintaining progress on longer tasks, trying alternatives and recovering from errors. These are descriptions of the intended agent behaviour; they should not be confused with the amount of text that fits into a single model call.[2]

Consider a hypothetical document assistant. Loading a long case file, recalling a preference from last week, and continuing a failed document conversion are three different jobs. A product may combine them, but the context window alone supplies only the first ingredient. This distinction helps explain why the surrounding application matters alongside the model.

A concrete example of the API bill

Artificial Analysis lists Wonder at $4 per million input tokens, $0.20 per million cached-input tokens and $20 per million output tokens. Using those rates, an illustrative request with 100,000 uncached input tokens and 5,000 billable output tokens costs $0.50. This is arithmetic, not a measured Muse session.[5]

If 80,000 of those input tokens qualify for the cached rate, the same calculation becomes $0.016 for cached input, $0.08 for the remaining input and $0.10 for output: $0.196 altogether. The saving in this example is 60.8%. Actual cache eligibility and all billable output, including reasoning where charged, have to be reflected in the usage record.[5]

Artificial Analysis’s own benchmark workload cost $4.52 per task for Wonder, compared with $1.60 for Spark 1.3—about 2.83 times as much. That workload-specific figure is more demanding than the small illustrative request above. It shows why a token price and the cost of completing an agent task answer different questions.[4]

For an API builder, the immediate opportunity is therefore concrete: a stronger Meta model that can be tried in the existing application. Full represents a separate product decision. Choose between them on the work to be completed and the access route available, rather than treating the family name as a single interchangeable subscription.

Report an error or suggest a correction