Meta · AI model
Muse Wonder Full
Variant of Meta's Muse Wonder series that Meta calls its most capable, announced 12 Oct 2026. Available only with a Muse Code subscription (Power Usage or Maximum Usage plan), not on the Meta Model API. Text, image, video and audio input, text output, 1M-token context.
Last verified Available 4 sources
Key specs
- Provider
- Meta
- Family
- Muse Wonder
- Availability
- Available
- Release date
- 2026-10-12
- License
- Proprietary
- Context window
- 1,000,000 tokens
- API access
- No
- Consumer app
- Yes
Capabilities
Inputs
- Text input Yes
- Image input Yes
- Audio input Yes
- Video input Yes
- File input Unknown
Outputs
- Text output Yes
- Image output No Output modality listed as text only
- Audio output No Output modality listed as text only
Features
- Reasoning Native Meta reports results at max reasoning effort (SWE-Marathon at xhigh)
- Tool calling Yes The model card describes multi-step workflows through tool use and interaction with applications; used as an agent in Muse Code
- Web search Unknown
"Unknown" means we have not verified this capability yet; it does not mean the model lacks it.
Pricing
No current prices listed.
Benchmarks
| Benchmark | Setting | Score | Reported by | Measured | Verified | Source |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 % resolved (pass@1), higher is better | v1.1 (113 tasks), pass@1; max effort; result Meta attributes to Datacurve, harness not stated (not on the public Datacurve board as of 2026-10-25) | 79 % | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) |
| GDPval-AA v2.1 Elo, higher is better | v2.1; max effort; Artificial Analysis Stirrup harness | 1,857 Elo | Third party | 2026-10-25 | 2026-10-25 | GDPval-AA v2.1 Leaderboard (Artificial Analysis) |
| OSWorld 2.0 % tasks completed, higher is better | v2.1, full 108-task set, binary (strict completion); max effort; GUI computer-control tool in Meta's internal evaluation framework | 39.4 % | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) |
| v2.1, full 108-task set, partial (mean per-task score, primary metric); max effort; GUI computer-control tool in Meta's internal evaluation framework | 78.4 % | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) | |
| Terminal-Bench 4.0 % tasks resolved, higher is better | Terminal-Bench 4.0; max effort; result sourced by Meta from Artificial Analysis (Artificial Analysis harness) | 57.1 % | Self-reported | — | 2026-10-25 | Muse Wonder Model Card (PDF) |
The official leaderboard uses mini-swe-agent for all models and shows whole-number percentages with margins of about ±1 to ±6 points. Providers often report scores from their own scaffolds or effort settings; DeepSeek reports 65.5% to 74.2% for the same model depending on the scaffold. DeepSWE v1 scores are not comparable with v1.1. With only 113 tasks, small differences are within noise.
Elo is a relative rating: values change as models are added and the scale is re-anchored, so always store the date a rating was read. For example, Anthropic's launch page gives 1846 for Claude Opus 5.5, while the leaderboard on 2026-10-12 shows 1866 (Max, Default Fallback). Earlier versions (Google cites "GDPVal-AA v2") are not comparable. Judging is done by an LLM, not by human experts. Anthropic models are listed with "Default Fallback", meaning refused steps are routed to a fallback model. AA-Briefcase and GDP.pdf are separate evaluations.
Numbers are only comparable when release, task set (full or offline), metric (binary or partial), step budget (150, 300 or 500) and agent approach (standard or batch tool calls) all match. Providers usually quote partial scores, which are much higher than binary completion. Claude Opus 5 at max effort on v2.1 scores 44.33% binary and 77.67% partial. The verified leaderboard is run by the maintainers and covers few current models, so most numbers are self-reported. The original OSWorld and OSWorld-Verified (369 tasks) are a different benchmark and are not recorded here.
Scores depend heavily on the agent harness (for example Claude Code, Codex, Grok Build or mini-SWE-agent) and on the reasoning effort, so the same model appears several times on the leaderboard. Leaderboard rows are often submitted by the organisation that builds the agent. Providers' own runs can differ from the leaderboard: Anthropic reports 66.4% for Claude Opus 5.5 at max effort, while the leaderboard lists 64.8% ± 3.1%. Versions 2.0, 2.1 and 3.0 and Artificial Analysis' "Terminal-Bench Hard" use different task sets and are not comparable with 4.0. Margins are about ±3 points.
One row per result, without rank numbers: each result was measured under its own setting (effort level, harness, tools, test version), as stated by the source, and scores compare only within the same setting. Self-reported results come from the provider and have not been reproduced independently.
Compare Muse Wonder Full
Guides
- AI token pricing explained: input, output, cached and batch rates
- Reasoning models explained: thinking modes, effort and cost
- Multimodal AI explained: image, audio, video and file inputs, and what they cost
- Context windows explained: limits, long-context pricing and recall
- How to choose an AI model: a practical decision framework
About Muse Wonder Full
Muse Wonder Full was announced on 12 October 2026 together with Muse Wonder. Meta calls it its most capable variant and ran its risk assessments on it. It is available only to Muse Code subscribers on the Power Usage ($50 per month) and Maximum Usage ($200 per month) plans; the Meta Model API lists no price for it. Meta's model card gives text, image, video and audio input, text output and a 1M-token context at inference for the Muse Wonder models.
Strengths and weaknesses
Strengths
- Meta's model card reports higher scores than Muse Wonder on every row except MemoryBench (for example DeepSWE v1.1 79.0 vs 78.8 and OSWorld V2.1 78.4 vs 72.6 partial, at max effort)
- 1M-token context at inference
- Text, image, video and audio input
Weaknesses
- Not available on the Meta Model API; only through a Muse Code subscription
- No per-token price; access is tied to the $50 and $200 monthly Muse Code plans
- Max output tokens and knowledge cutoff are not published
Sources
Last verified Page updated 2026-10-26