Muse Wonder
Released Oct 12, 2026 · knowledge cutoff unpublished
- Status
- Frontier
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- —
- Speed (max effort)
- 85 tok/s · 61s to first answer token
- Price ($/MTok in / out)
- $4 / $20
- Cost per task (max effort)
- $4.52
- Open weights
- No
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Carries what it learns about you across conversations and pulls it back in when a later task needs it — memory is trained behaviour rather than a bolt-on retrieval layer
- Keeps working through blockers on long agent runs, trying another route instead of stalling, and checks in before important actions
- Reasoning effort runs from minimal to max on one model id, so quick lookups and multi-hour agent work share the same endpoint
- Makes up answers less often than Muse Spark 1.3 when it doesn't know
Weaknesses
- Long silence before the first answer token at max effort — the reasoning phase runs about a minute
- Visual and spatial reasoning, especially 3D, is a limitation Meta's own model card tells developers to validate
- Still a preview with no hands-on reception yet, and every launch figure was run at max effort — behaviour at the cheaper settings is undocumented
Preview on the Meta Model API (muse-wonder-1.0, Standard mode only, no Contributor tier) and in Muse Code; cached input is $0.20/M and there is no long-context premium. Meta published none of the tracked benchmarks: its launch table (all max effort, with competitor runs Meta itself calls best-effort) reports its own suite — Terminal-Bench 4.0-AA 52.0% (not the tracked 2.1), DeepSWE 1.1 78.8%, SWE-Marathon 53.1%, ProgramBench 50.0%, MemoryBench 85.6%, Personal Superintelligence Index 54.3. Artificial Analysis scores it 52 on Intelligence Index v4.3 at max effort (the only variant it lists); AA's HLE 50.1% comes from its own text-only subset with an unstated tool setting and its GDPval-AA 1770 is v2.1, a different scale from the tracked v2, so hle and gdpvalAA stay null. costPerTask and speed are AA's max-effort measurements on Meta's first-party endpoint. maxOutput is unpublished by Meta (OpenRouter's 943,718 max-completion figure is not official); knowledge cutoff undisclosed. Coverage disagreed on the comparison set: TechCrunch named only Claude Fable 5.1 and GPT-6 Astra, but Meta's model card also compares Claude Opus 5.5, which leads Wonder on AA's index. predecessorId is null: Meta says the release 'builds on' Muse Spark but names no model it replaces, and its docs keep recommending Muse Spark for latency- and cost-sensitive work.
For developers
API model strings
- meta
muse-wonder-1.0 - openrouter
meta/muse-wonder
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor