Muse Wonder is Meta's frontier agentic model, released on 12 October 2026 and presented by Meta as its most capable model to date. It is aimed at work that runs long rather than work that answers fast: agents that plan a task, delegate parts of it to sub-agents, call tools in parallel and recover from their own errors instead of stopping at the first failure.
Meta ships it in two variants. Muse Wonder, which Meta serves through its own Model API as muse-wonder-1.0, is the one available here as meta/muse-wonder. Muse Wonder Full (muse-wonder-full-1.0) is the larger variant and is sold only inside Muse Code subscriptions, so it is not an API model.
Meta names four headline capabilities: frontier coding across large repositories, with multi-file changes made and verified in parallel; long-running agents that plan work, delegate to sub-agents, call tools in parallel and recover from errors; advanced computer use, where an agent sees and operates browsers, desktop and mobile apps; and active memory that carries across conversations.
Meta's scorecard sets Muse Wonder and Muse Wonder Full against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, all at max reasoning effort. Muse Wonder's strongest published figures are MemoryBench 85.6 and DeepSWE 1.1 at 78.8, both ahead of all three compared models, and SWE-Marathon at 53.1. It is reported behind Claude Opus 5.5 on Terminal-Bench 4.0-AA (52.0 against 59.6), on ProgramBench (50.0 against 70.5) and on HealthBench Pro (60.5 against 65.6). On Artificial Analysis's Intelligence Index v4.3 Meta reports 52 for Muse Wonder and 56 for Muse Wonder Full, against 58 for Claude Opus 5.5.
Meta's own methodology notes that some rows use third-party result sources while others are internal, that one coding row is not a controlled common-harness comparison, and that competitor runs may not be tuned to those vendors' strengths. None of the figures has been reproduced here.
Meta's model card names visual and spatial reasoning, especially 3D tasks, as a known limitation, and notes the model has not been evaluated on every language in its pre-training data. The card also records that, without mitigations, Meta does not rule out the model reaching the high-risk threshold in the chemical and biological domain under its Superintelligence Scaling Framework, which is why domain-specific safety post-training and system-level classifiers are applied.
Sources: dev.meta.ai, research.meta.ai
Meta documents Muse Wonder for frontier coding across large repositories, where the model makes and verifies multi-file changes in parallel rather than editing one file at a time.
It is also documented for long-running agents: work that is planned, delegated to sub-agents, carried out with tools called in parallel, and resumed after an error rather than abandoned.
Meta lists advanced computer use — agents that see and operate browsers, desktop and mobile applications — and active memory that carries context from one conversation into the next.
In Meta's own framing the model is built for personal and professional agents handling multi-step real-world tasks, and for domain knowledge work in law, finance and health: its published scorecard includes legal, financial-analyst and professional health evaluations alongside the coding ones.
import requests r = requests.post( "https://api.aimlapi.com/v1/chat/completions", headers={"Authorization": "Bearer " + AIMLAPI_KEY}, json={ "model": "meta/muse-wonder", "messages": [ { "role": "user", "content": "Hello!" } ] }, ) print(r.json())
const r = await fetch("https://api.aimlapi.com/v1/chat/completions", { method: "POST", headers: { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }, body: JSON.stringify({ "model": "meta/muse-wonder", "messages": [ { "role": "user", "content": "Hello!" } ] }), }); console.log(await r.json());
curl -X POST https://api.aimlapi.com/v1/chat/completions \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"meta/muse-wonder","messages":[{"role":"user","content":"Hello!"}]}'
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Cached input |
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| Personal Superintelligence Index | 54.3 | Composite of seven personal and knowledge-work evaluations, reported by Meta | Source | October 12, 2026 |
| Harvey LAB-AA | 10.0 | Legal work products built from matter documents, graded against task-specific criteria with hallucination checks | Source | October 12, 2026 |
| Finance Agent v2 | 56.6 | Financial research and analysis run as an agent | Source | October 12, 2026 |
| HealthBench Professional | 60.5 | Clinical and professional health question answering | Source | October 12, 2026 |
| AutomationBench | 39.1 | Completion of routine automation workflows | Source | October 12, 2026 |
| GDP.pdf | 30.8 | Professional work grounded in supplied PDF source material | Source | October 12, 2026 |
| JobBench | 70.5 | 65 professional tasks across 35 white-collar occupations, each scored against task-specific rubrics | Source | October 12, 2026 |
| OSWorld | 72.6 | Computer-use across real desktop applications | Source | October 12, 2026 |
| MemoryBench | 85.6 | Recall of critical details across long multi-turn conversations with distractors and later updates | Source | October 12, 2026 |
| Terminal-Bench 4 | 52.0 | Autonomous shell and terminal task completion, fourth release | Source | October 12, 2026 |
| DeepSWE | 78.8 | Agentic software engineering: resolving real repository tasks end to end | Source | October 12, 2026 |
| SWE-Marathon | 53.1 | Ultra-long-horizon software engineering on substantial projects, graded by executing project tests | Source | October 12, 2026 |
| ProgramBench | 50.0 | Rebuilding a complete program in a clean room, graded by a hidden behavioural test suite | Source | October 12, 2026 |
| Intelligence | 52.4 | Composite score across standardised reasoning, knowledge and problem-solving evaluations, measured independently by Artificial Analysis | Source | October 12, 2026 |
Figures reported by Meta at release, at max reasoning effort; not independently reproduced.
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Muse Wonder This page | Long-running agents, computer use and large-repo coding | |||
| Agentic workflows and structured output | ||||
| Long-context multimodal reasoning and tool use | ||||
| Multimodal reasoning and tool use |
Yes. Muse Wonder is live on AI/ML API as meta/muse-wonder, through the chat completions endpoint.
Muse Wonder has a 1,048,576 tokens context window and can return up to 943,718 tokens.
Muse Wonder takes image, text and video as input and returns text. Meta's own model card also lists audio input.
Muse Wonder is the variant Meta serves through its Model API. Muse Wonder Full is the larger variant and is available only to Muse Code subscribers on Meta's Power Usage and Maximum Usage plans, so it is not offered as an API model. Meta reports Muse Wonder Full ahead of Muse Wonder on almost every published benchmark.
Meta positions it for long-running agents, computer use and coding across large repositories, with memory that carries across conversations. Its strongest published figures are MemoryBench 85.6 and DeepSWE 1.1 at 78.8, both ahead of GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 in Meta's scorecard.
Muse Wonder is priced at input $5.5016 / 1M tokens, output $27.508 / 1M tokens, cached input $0.27508 / 1M tokens.
No. The figures on this page come from Meta's own scorecard published at release, with the model at max reasoning effort. Meta's methodology notes that some rows use third-party result sources while others are internal, and that competitor runs may not be tuned to those vendors' strengths.
