What are Meta's Muse Wonder AI models?
Muse Wonder and Muse Wonder Full are Meta's frontier AI models. Features, benchmark scores, pricing, API setup and limits as of 12 October 2026.
Muse Wonder and Muse Wonder Full are Meta's frontier models for personal and professional AI agents, released on Monday, 12 October 2026. According to Meta Platforms research from the Meta AI research blog, 12 October 2026, Muse Wonder and Muse Wonder Full are Meta's "most capable models that deliver state-of-the-art performance on personal and professional agentic tasks." Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is available to Muse Code subscribers.
According to Meta Platforms research from the Muse Wonder Model Card, 12 October 2026, Muse Wonder and Muse Wonder Full accept text, image, video and audio input, output text, and run with a context length of 1 million tokens at inference.
Key takeaways
- Identify the features: Muse Wonder accepts text, image, video and audio input and runs with a 1 million token context length.
- Compare the scores: Muse Wonder Full scores 56 points on the Artificial Analysis Intelligence Index v4.3.2, 2 points behind Claude Opus 5.5.
- Check the costs: Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens.
- Follow the setup: The Meta Model API serves muse-wonder-1.0 at a limit of 3,000 requests per minute per team.
- Assess the limits: The Muse Wonder Model Card lists 6 limitations, including errors in multi-step reasoning in novel scenarios.
Meta Platforms as model provider
Meta Platforms is the model provider of Muse Wonder and Muse Wonder Full, listed as Meta Platforms, Inc. or Meta Platforms Ireland Limited. Meta trained Muse Wonder and Muse Wonder Full at its gigawatt-scale Prometheus data center, across pretraining and reinforcement learning, and assessed Muse Wonder and Muse Wonder Full under the Meta Superintelligence Scaling Framework.
Muse model family
According to Meta Platforms documentation from the Meta Model API docs, 16 October 2026, the Muse family has 9 entries, listed in the table with what each entry is and how each entry is accessed.
Entry | What it is | Access |
|---|---|---|
Muse Wonder | Model ID muse-wonder-1.0, marked "in preview" | Meta Model API and Muse Code |
Muse Wonder Full | Most capable variant of the Muse Wonder pair | Muse Code subscribers |
Muse Spark | Model IDs muse-spark-1.3, muse-spark-1.2 and muse-spark-1.1 | Meta Model API, with Contributor variants for 1.3 and 1.2 |
Muse Spark Open | 816B-parameter mixture-of-experts model with 42B active parameters | Self-hosted or through hosted providers, as BF16 or NVFP4 checkpoints |
Muse Glimmer | Open-weight model | Self-hosted, Apache 2.0 license |
Muse Image | Model ID muse-image-1.0, text-to-image and editing | Meta Model API |
Muse Voice Transcribe | Model ID muse-voice-transcribe-1.0, speech-to-text only | Meta Model API |
SAM 3.1 | Model ID sam-3.1, with text, image and video input and box and mask output | Meta Model API |
Muse Code | Command-line coding agent | Install script, then browser sign-in or API key |
The Meta Model API docs list muse-image-1.0 for text-to-image and editing. The docs list Muse Spark with its own model IDs and prices, separate from muse-wonder-1.0.
What features do the Muse Wonder models have?
The Muse Wonder models have 4 main features: Active Memory, Active Focus, multimodal input with a 1 million token context window, and tool use with sub-agent delegation. Meta states that Muse Wonder and Muse Wonder Full deliver state-of-the-art results on key benchmarks for personal and professional agents, memory and coding.
Active Memory
Active Memory is carrying important context about the user across conversations, with active retrieval of task-relevant details. MemoryBench is Meta's internal benchmark that tests how accurately a model recalls critical information across multi-turn conversations of 64K to 100K tokens, with distractor conversations and injected turns that conflict with or later update the target details. An LLM judge scores each MemoryBench answer against weighted, question-specific rubric criteria.
On MemoryBench, Muse Wonder scores 85.6, Muse Wonder Full scores 85.5, Claude Opus 5.5 scores 81.5, GPT-6 Astra scores 81.4 and Claude Fable 5.1 scores 79.0, in Meta's figures. Data permissioning limits data liquidity, and memory that Muse Wonder keeps across conversations raises the same permissioning questions as any reused data asset. The article Data Liquidity: How Easily Can Your Data Be Reused and Combined? explains how data permissioning affects reuse.
Active Focus
Active Focus is how Muse Wonder continues working toward a challenging goal: Muse Wonder tries alternative approaches, works through blockers, adapts to new information, asks for clarification and checks in before important actions. An agent that retries a blocked step with a different approach and then asks before an important action shows Active Focus.
FocusBench is Meta's internal evaluation of personal agentic tasks across domains such as housing, travel, IT and finance. The Muse Wonder Model Card gives no FocusBench score.
Multimodal input and context window
Muse Wonder accepts text, image, video and audio input, outputs text, and runs with a context length of 1 million tokens at inference. The Meta Model API docs list the context window of muse-wonder-1.0 as 1,048,576 tokens. Muse Wonder outputs text only.
A longer prompt costs more input tokens, at $4.00 per million input tokens. A Sparse Priming Representation supplies the idea in fewer tokens than the raw document, and the article Sparse Priming Representations: Give a Language Model the Idea, Not the Document explains how.
Tool use and sub-agents
Muse Wonder is built for complex software engineering and long-horizon agentic work, and supports parallel tool calls and sub-agent delegation. Muse Wonder powers experiences in Muse, rolling out first to subscribers, and artifact generation.
The Muse Wonder Model Card lists personal and professional agents as an intended use, supporting research, planning and multi-step workflows through tool use and interaction with applications. The Meta Model API docs mark Muse Wonder as in preview.
How do the Muse Wonder models perform on benchmarks?
According to Artificial Analysis research from the Artificial Analysis Intelligence Index v4.3.2, 16 October 2026, Muse Wonder Full scores 56 and Muse Wonder scores 52, against 58 for Claude Opus 5.5. Meta's scorecard reports 14 benchmark rows for 5 models, and Meta states that the comparison mixes controlled internal runs with third-party and provider-reported results.
Artificial Analysis Intelligence Index
The table lists the Artificial Analysis Intelligence Index v4.3.2 scores of 6 models at Max reasoning effort. The index combines 10 evaluations.
Model | Intelligence Index v4.3.2 score |
|---|---|
Claude Opus 5.5 | 58 |
Claude Sonnet 5.5 | 56 |
Muse Wonder Full | 56 |
GPT-6 Astra | 53 |
Claude Fable 5.1 | 53 |
Muse Wonder | 52 |
Muse Wonder Full is 2 points behind Claude Opus 5.5 and 3 points ahead of GPT-6 Astra. Muse Wonder is 1 point behind GPT-6 Astra and 6 points behind Claude Opus 5.5.
Meta benchmark scorecard
The table lists 14 benchmark rows with Meta's scores for 5 models, all at max reasoning effort.
Benchmark | What it measures | Muse Wonder Full | Muse Wonder | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|---|
Personal Superintelligence Index | Meta's internal composite of 7 personal and knowledge-work evaluations | 59.1 | 54.3 | 49.9 | 54.5 | 55.2 |
Harvey LAB-AA | Legal agent work | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
Vals Finance Agent V2 | Financial analyst research | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
HealthBench Pro | Open-ended health queries | 62.3 | 60.5 | 62.1 | 65.3 | 65.6 |
AutomationBench-AA | End-to-end business workflows | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
GDP.pdf | Professional work | 33.2 | 30.8 | 27.6 | 34.2 | 30.6 |
JobBench | Professional tool use | 75.3 | 70.5 | 73.1 | 66.1 | 75.4 |
OSWorld V2.1 (partial) | Agentic computer use | 78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
OSWorld V2.1 (binary) | Agentic computer use | 39.4 | 34.3 | 43.1 | 34.3 | no figure |
MemoryBench | Remembering context | 85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
Terminal-Bench 4.0-AA | Agentic terminal coding | 57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
DeepSWE 1.1 | Long-horizon coding | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
SWE-Marathon | Ultra-long-horizon engineering | 57.5 (xhigh) | 53.1 | 45.6 | 42.5 | 46.3 |
ProgramBench | From-scratch programming | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
Muse Wonder Full scores highest on 6 of the 14 rows: Personal Superintelligence Index, Harvey LAB-AA, Vals Finance Agent V2, AutomationBench-AA, DeepSWE 1.1 and SWE-Marathon. Claude Opus 5.5 scores highest on 5 rows: HealthBench Pro, JobBench, OSWorld V2.1 (partial), Terminal-Bench 4.0-AA and ProgramBench. Muse Wonder scores highest on MemoryBench, GPT-6 Astra scores highest on GDP.pdf, and Claude Fable 5.1 scores highest on OSWorld V2.1 (binary).
On Harvey LAB-AA, the legal work row, Muse Wonder Full scores 13.1 against 4.2 for Claude Opus 5.5. On AutomationBench-AA, the business workflow row, Muse Wonder Full scores 45.4 against 41.2 for Claude Opus 5.5.
Evaluation conditions
Meta's scorecard follows 4 evaluation conditions:
- Run results at max reasoning effort, with Muse Wonder Full at xhigh reasoning effort on SWE-Marathon.
- Select the comparable third-party or provider-reported result where one exists, and otherwise the highest of Meta's internally validated run, self-reported score or official leaderboard score.
- Score model refusals as zero credit, with refusals kept in the denominator.
- Compare SWE-Marathon without a common harness, because GPT-6 Astra runs in Codex and Claude Fable 5.1 runs in Claude Code.
Meta's internally generated competitor runs are best-effort and not tuned for proprietary models. DeepSWE uses third-party results from Datacurve for Muse Wonder, Muse Wonder Full and Claude Opus 5.5, and the Claude Opus 5.5 result on OSWorld V2.1 is Anthropic's provider-reported partial score with no binary figure.
Muse Wonder vs Claude Opus 5.5 and GPT-6 Astra
The table lists 9 measures from the Artificial Analysis pages for Muse Wonder and Claude Opus 5.5, both at Max reasoning effort.
Measure | Muse Wonder | Claude Opus 5.5 |
|---|---|---|
Intelligence Index score | 52 points | 58 points |
Input price | $4.00 per million tokens | $4.00 per million tokens |
Output price | $20.00 per million tokens | $20.00 per million tokens |
Blended price (7:2:1 cache hit, input, output ratio) | $2.94 per million tokens | $2.94 per million tokens |
Output speed | 185.5 tokens per second | 96.3 tokens per second |
Time to first token | 36.32 seconds | 666.74 seconds |
Cost per Intelligence Index task | $4.52 | $5.98 |
Cost to run the full Intelligence Index | $5,733 | $8,708 |
Context window | 1049k tokens | 1000k tokens |
GPT-6 Astra, developed by OpenAI, costs $10.00 per million input tokens and $50.00 per million output tokens and scores 53 points, against $4.00, $20.00 and 52 points for Muse Wonder. Muse Wonder used 130M output tokens on the Intelligence Index against a median of 87M.
How much do the Muse Wonder models cost?
Muse Wonder costs $4.00 per million input tokens and $20.00 per million output tokens through the Meta Model API. Cached input costs $0.20 per million tokens, and web search grounding costs $2.50 per 1,000 queries on top of token costs. Muse Wonder Full is available to Muse Code subscribers, and the Meta Model API docs list no per-token price for Muse Wonder Full.
API pricing
The table lists the prices of 4 model and mode combinations in US dollars per million tokens, at launch.
Model ID | Mode | Cached input | Input | Output |
|---|---|---|---|---|
muse-wonder-1.0 | Standard | 0.20 | 4.00 | 20.00 |
muse-wonder-1.0 | Contributor | not available | not available | not available |
muse-spark-1.3 | Standard | 0.15 | 1.25 | 4.25 |
muse-spark-1.3-contributor | Contributor | 0.002 | 0.10 | 0.20 |
Muse Spark 1.1 is available in Standard mode only.
Rate limits
The table lists the rate limits of 3 model IDs, applied per team.
Model ID | Requests per minute | Tokens per minute |
|---|---|---|
muse-wonder-1.0 | 3,000 | 4,000,000 |
muse-spark-1.3-contributor | 150 | 3,000,000 |
muse-spark-1.2-contributor | 150 | 3,000,000 |
Rate limits apply per team, not per API key. Background response submissions are capped at 600 per minute per team.
Free access
Muse Wonder is priced per token on the Meta Model API. Muse Wonder Full is available to Muse Code subscribers.
How do you use the Muse Wonder models?
To use the Muse Wonder models, send requests to the Meta Model API with the model ID muse-wonder-1.0, or run Muse Code. The Meta Model API is drop-in compatible with the OpenAI SDK, the Anthropic SDK and OpenAI-compatible agent CLIs, and the docs mark Muse Wonder as in preview.
Meta Model API setup
The Meta Model API takes 4 steps to send a first request to Muse Wonder.
- Set the base URL to https://api.meta.ai/v1 for every request.
- Send the bearer token in the Authorization header, as the docs examples do with MODEL_API_KEY.
- Choose the protocol: the Responses API at POST /v1/responses, the Chat Completions API or the Messages API.
- Call Muse Wonder with the model ID muse-wonder-1.0.
The 4 steps cover the request path that the Meta Model API docs state.
Muse Code
Muse Code is the command-line coding agent. The install command for macOS or Linux is:
curl -fsSL https://dev.meta.ai/install.sh | sh
The install command for Windows is:
irm https://dev.meta.ai/install.ps1 | iex
Each install command downloads and runs a script from dev.meta.ai. The command muse starts Muse Code, and on first run Muse Code offers a browser sign-in or an API key.
Intended uses
The Muse Wonder Model Card lists 6 intended uses, in this order:
- General-purpose assistance and reasoning
- Coding and software engineering
- Personal and professional agents
- Domain-specific knowledge work, such as law, finance and health
- Multimodal understanding and reasoning
- Research and development
The Muse Wonder Model Card states that Muse Wonder and Muse Wonder Full are intended for commercial and research use.
What are the limits and risks of the Muse Wonder models?
The Muse Wonder models have 3 main limits and risks: limitations in the model card, risk assessment results and data handling. Meta states that Muse Wonder carries known and unknown risks, and that testing to date has not covered all scenarios.
Model card limitations
The Muse Wonder Model Card lists 6 limitations.
- Produce inaccurate, biased or objectionable responses to user prompts.
- Make errors in multi-step reasoning, particularly in novel scenarios not represented in training data.
- Trail on visual and spatial reasoning, especially 3D tasks.
- Degrade outside the strongly supported languages, because not all pre-training languages were evaluated.
- Vary with quantization, serving configuration, prompts, tools and the agent harness.
- Exclude individuals under 18, with deployers responsible for assessing and mitigating the risks of such use.
Risk assessment results
The risk assessment is Meta's evaluation of Muse Wonder Full under the Meta Superintelligence Scaling Framework, and the assessment covers every model in the family. Without mitigations, Meta does not rule out Muse Wonder Full reaching the high risk threshold in the Chemical and Biological domain. Meta's evaluations put the unmitigated Muse Wonder Full at "moderate or lower" for Cybersecurity and Loss of Control. With mitigations and system-level safeguards, Meta reaches "moderate or lower" across all domains.
The table lists 7 evaluations with the results of Muse Wonder Full, all from Meta's own evaluations.
Evaluation | Muse Wonder Full result |
|---|---|
CyberGym | 88.4% |
ExploitBench | 58.9% |
ExploitGym | 92 flags captured, 21 through the intended vulnerability |
BioTIER refusal | 98.7% English, 98.7% multilingual, 98.2% adversarial |
Internal chemistry refusal | 100.0% English, 100.0% multilingual, 99.8% adversarial |
MASK honesty | 87.3% (higher is better) |
Agentic misalignment | 0% (lower is better) |
Data handling
The Meta Model API docs price muse-wonder-1.0 in Standard mode, and Contributor mode is not available for muse-wonder-1.0. Meta lists a full safety report with additional evaluations as available soon.
Meta recommends deploying Muse Wonder as part of a broader AI system with guardrails suited to the use case, such as human-in-the-loop confirmation for irreversible actions, and recommends that developers evaluate Muse Wonder for their own deployment and languages. A system is observable when failures are seen before users report them, and an agent that acts on a business system is covered by the same test, as the article A System That Lasts: Understandable × Observable × Recoverable describes.
To prepare data and systems for AI agents, read about Information Systems or Contact me.