Muse Wonder

Meta's most capable model, built for autonomous agentic coding and long-horizon software work.

Context1M

ModalityText · Image · Video · Audio

AccessAPI

ReasoningControllable

Where it fits.

Pairs Active Memory and Active Focus to hold multi-session context, self-correct tool loops and confirm before irreversible actions. Released October 12, 2026 as muse-wonder-1.0 at roughly $4 input and $20 output per million tokens on a 1M-token context, with $0.20 cached input. It accepts text, image, video, audio and PDF input, emits text only, and supports reasoning effort from minimal to max plus computer use and search grounding.

Core capabilities

1M-token context with text output only
Active Memory and Active Focus for long agent runs
Reasoning effort from minimal through max
Computer use, structured output and search grounding

Best-fit workflows

Autonomous coding agents
Multi-file refactoring
Parallel tool workflows

04 / OPERATIONS

From evaluation
to production.

Deployment notes

Adopt for agentic coding only with preview-level expectations. Put invariant constraints at the start of the system prompt, state objectives rather than step sequences, set reasoning effort explicitly, and keep Spark 1.3 as the supported fallback.

Watch before choosing

Still labelled preview, and the stronger Wonder Full is available only inside Muse Code subscriptions rather than the public API. Meta lists no benchmark figures and marks both Wonder and Wonder Full as preview. Wonder Full is gated behind $50-200 per month Muse Code plans; EU availability and data residency are unconfirmed.

Family & versions

Muse WonderFrontier flagship
Muse Spark OpenOpen weights
Muse Spark 1.3Recommended API tier
Llama 4 MaverickOpen line

Specifications and positioning checked against provider documentation.