Muse Wonder Full
Released Oct 12, 2026 · knowledge cutoff unpublished
- Status
- Frontier
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- 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
- Holds up better on unfamiliar tasks than its post-training targets would predict — Meta credits a larger base model rather than benchmark tuning
- Gets more done per token than the standard Wonder, so the stronger model is also the less wasteful one on long agent runs
- Built for multi-hour coding inside Muse Code, sustaining whole-project work rather than single-file edits
Weaknesses
- No API at all — reachable only inside Muse Code on the top two subscription plans, so it can't be built into your own product
- Ships behind CBRN classifiers after Meta couldn't rule out the unmitigated model reaching its high-risk threshold, so dual-use biology requests draw heavy refusals
- Too new for hands-on reception — beyond Meta's launch materials its day-to-day behaviour is undocumented
Subscription-only: Muse Code's Power Usage ($50/month) and Maximum Usage ($200/month) plans, in Standard or Contributor mode, so pricing is null — there is no per-token rate. Meta's pricing table lists the string muse-wonder-full-1.0 but it is not served on the Meta Model API (setting max effort on Muse Wonder does not select it), so apiIds is empty and availability is restricted; CNBC's $20/$100 monthly tiers are for the separate Muse consumer agent app. Meta states only '1M tokens' of context (its docs give 1,048,576 for Muse Wonder but no exact Full figure). Meta published none of the tracked benchmarks: its launch table (max effort, competitor runs best-effort) reports Terminal-Bench 4.0-AA 57.1% (not the tracked 2.1), DeepSWE 1.1 79.0%, SWE-Marathon 57.5% (at xhigh), ProgramBench 59.0%, Harvey LAB-AA 13.1% and Personal Superintelligence Index 59.1. Artificial Analysis scores it 56 on Intelligence Index v4.3 at max effort, behind Claude Opus 5.5's 58; its HLE 53.6% (own text-only subset, tool setting unstated) and GDPval-AA v2.1 1857 (not the tracked v2) are left null, and AA publishes no cost per task (not priced per token) or speed (no hosted endpoint). maxOutput and knowledge cutoff are unpublished; predecessorId is null because Meta names no model it replaces.
For developers
API model strings
Not researched
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor