Meta’s Muse Wonder is stronger—but its best version isn’t on the API

Two models, different doors

Wonder is available as an API preview and in Muse Code, Meta’s coding agent. Full is available only through Muse Code’s Power Usage and Maximum Usage plans. Turning standard Wonder’s reasoning setting to “max” does not turn it into Full; these are separate variants.

Meta’s launch post also says Wonder will power experiences in the Muse personal agent later, beginning with subscribers. That is a promised rollout, not confirmation that everyone using Muse already has the new model.

Remember the context, keep working

Meta’s two headline features describe familiar problems with agents. Active Memory is meant to retain relevant details across conversations and retrieve them when needed. Active Focus is meant to keep a task moving through obstacles: try another approach, recover from an error and ask for clarification rather than quietly stop.

The company demonstrates tasks such as arranging a wedding seating chart and designing custom gifts from Instagram photos. These are vendor demonstrations, not independent measures of reliability. A large context window—the amount a model can receive in one request—is also different from a dependable memory that survives across conversations.

The improvement survives an outside test

Artificial Analysis measured standard Wonder at 52 and Full at 56 on its Intelligence Index, compared with 48 for Muse Spark 1.3. All three used maximum reasoning effort. This is a composite benchmark score, not a percentage of everyday jobs completed correctly.

The evaluator reports Full improving across all ten component tests. It also flags knowledge as a relative weakness despite comparatively low hallucination rates. The result is meaningful progress, with uneven strengths; it does not establish that either model will be best for every task.

Meta’s own model card is useful for understanding the wider launch scorecard. It mixes internal runs, outside evaluations and provider-reported results, and acknowledges that some competitor configurations may not be optimally tuned. Those rows should not be read as one perfectly controlled race.

Fewer output tokens can still mean a bigger bill

Standard Wonder used fewer output tokens than Spark in Artificial Analysis’s tests. Yet its higher prices and greater input use pushed the weighted cost to $4.52 per Intelligence Index task, against $1.60 for Spark—about 2.8 times as much. Input processing, including cache writes and reads, accounted for most of Wonder’s cost.

The comparison below is a measured benchmark bill, not a quote for answering one email or completing one campaign. It includes each tested configuration’s token use and prices. Full has no published per-token price, so Artificial Analysis does not provide a comparable task-cost figure for it.

Small tasks and a rocky first night

A smaller test offers a useful contrast. Piotr Olszewski, publisher of the Polish AI site Promptowy, ran 23 tasks through OpenRouter at low reasoning effort, with automatic checks rather than an AI judge. Wonder scored 99.7% and cost $0.64 for the suite; Opus 5.5 scored 99.4% and cost $0.75. Wonder counted the same long Polish document in fewer tokens.

Those near-perfect scores leave little room to separate the models. It was one run per task, not a long-running agent trial, and Olszewski discloses that he also sells AI-tool accounts. His examples include a spelling error in generated invoice code and an invented date in a draft letter. The latter did not lose a point: even a high score can leave something a reader needs to catch.

Subscription users also encountered a launch issue. Meta’s chief AI officer Alexandr Wang said the company had corrected unusually fast quota consumption for Wonder and Full and reset Muse Code usage.

Wang’s edited post gives Meta’s account of the fix. It is not independent confirmation that every affected session now behaves correctly; reliable comparisons of subscription value will need to account for that correction.

A finished run is not always a finished job

Longer creative builds give a less flattering result. In his 12 October review, Pat Simmons, publisher of AI for Mortals, gave Wonder Full, Opus 5.5, Fable 5.1 and GPT-6 Astra the same four prompts: a website clone, a coffee launch campaign, an iPhone app and a 3D rocket simulation. Each ran at high reasoning effort in its own lab’s coding agent. Simmons published the prompts and outputs, and ranked Full last in all four.

That is one reviewer’s qualitative comparison across different agents, not a general coding leaderboard. His assessment of the campaign films is especially sobering. Readers can inspect Full’s campaign gallery alongside the brief; producing the requested files and producing usable creative work are different tests.

His cost table needs a separate caution. Full has no published API rate, so he priced its tokens using standard Wonder’s rates. That estimate is not Full’s subscription bill or evidence that Full is cheapest. His Muse runs hit a usage limit and resumed after an upgrade; paused time was excluded, and their timing relative to Meta’s quota correction is unresolved.

Access and data settings deserve the same attention

For European teams, model access and the consumer Muse rollout are separate questions. Meta’s developer geographic policy does not list EU countries among its restricted territories, but it says access depends on where Meta has enabled the service. That is not a blanket guarantee of availability, a European hosting commitment or a new launch date for Muse for Small Business.

Data settings are more concrete. Meta says Standard mode does not use submitted content to train its models. Muse Code subscriptions default to Contributor mode, which permits training use in return for more usage; users can turn it off. Pay-as-you-go Wonder is Standard-only. That distinction matters before submitting a client’s code, unpublished plans or private documents.

The useful next test is therefore specific: can the version you can actually access finish your task, under your data settings and usage limit? Meta has made a measurable step forward. Choosing it still requires looking past the largest score in the launch graphic.