Meta Superintelligence Labs · Released October 12, 2026 · Guide
Muse Wonder by Meta: benchmarks, pricing and how to use it
Everything about Meta's Muse Wonder on one page: what Active Memory and Active Focus do, the official benchmark table, API prices, a muse-wonder-1.0 quickstart and where you can use it today.
Muse Wonder can't make pictures
Need images? Try Nano Banana 2.1 on Nano AI
Muse Wonder reads images but only answers in text, and Meta's video model is not out yet. To turn its ideas into pictures, ask it for an image prompt and paste it here: Nano Banana 2.1 makes the image in your browser for 8 credits, with no API key, and GPT Image 2.5, Seedream 5.0 and other image models are in the same menu.
- Made by
- Meta
- Released
- October 12, 2026
- Run on Nano AI
- Nano Banana 2.1 · 8 credits per image
- Page updated
- October 20, 2026
Source: research.meta.ai
What is Muse Wonder?
Muse Wonder is a frontier AI model from Meta Superintelligence Labs, released on October 12, 2026. It belongs to Meta's Muse family, next to Muse Spark and Muse Image, and is not a Llama model. There are two versions: Muse Wonder, which developers can call through Meta's API, and Muse Wonder Full, Meta's most capable variant, which comes only with a Muse Code subscription.
Meta built it around two ideas. Active Memory carries what it knows about you across conversations and pulls up the details a task needs; Active Focus keeps it working towards a hard goal, trying other approaches when it gets stuck and checking in before important actions. It reads text, images, video, audio and PDFs within a 1-million-token context window, answers in text, and its weights are not public.
Muse Wonder does not create images or video. Meta's image model is the separate Muse Image, and Muse Video has been shown but not released. If you came here to make pictures, Nano AI runs Nano Banana 2.1 and other image models in the browser.
Muse Wonder specifications
- Developer
- Meta Superintelligence Labs
- Released
- October 12, 2026
- Versions
- Muse Wonder (API) and Muse Wonder Full (subscription only)
- API model ID
- muse-wonder-1.0, in preview
- Context window
- 1,048,576 tokens
- Input
- Text, images, video, audio and PDF
- Output
- Text only
- Reasoning effort
- minimal to max; cannot be turned off
- API price
- $4 input, $0.20 cached input, $20 output per 1M tokens
- Rate limit
- 3,000 requests and 4M tokens per minute per team
- Weights
- Closed
- Not available in
- South Korea
Muse Wonder API quickstart
Muse Wonder is in Meta's Model API as muse-wonder-1.0, at https://api.meta.ai/v1. It works with the OpenAI SDK (Responses and Chat Completions) and the Anthropic SDK (Messages).
- 1
Create an API key
Sign in at dev.meta.ai with a Meta account, open API keys in the dashboard and choose Create API key. Keep it in an environment variable.
export MODEL_API_KEY=your_key_here - 2
Send a first request
This is Meta's own sample request, with a reasoning effort set:
curl -X POST "https://api.meta.ai/v1/responses" \ -H "Authorization: Bearer $MODEL_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"muse-wonder-1.0","reasoning":{"effort":"high"},"input":"Summarise this contract in five bullet points."}' - 3
Or use the OpenAI SDK
Point the client at Meta's base URL. The SDK does not read MODEL_API_KEY by itself, so pass the key explicitly.
import os from openai import OpenAI client = OpenAI(base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"]) response = client.responses.create( model="muse-wonder-1.0", reasoning={"effort": "high"}, input="Plan a three-day trip to Kyoto.", ) print(response.output_text) - 4
Choose a reasoning effort
Effort can be minimal, low, medium, high, xhigh or max; leave it out and the model picks. Reasoning cannot be switched off (none returns HTTP 400), and max does not turn it into Muse Wonder Full.
- 5
Or code with Muse Code
Meta's command-line coding agent for macOS and Windows. Every plan includes Muse Wonder; Power and Maximum add Muse Wonder Full.
curl -fsSL https://dev.meta.ai/install.sh | bash
Where you can use Muse Wonder
Availability on October 20, 2026. Muse Wonder Full is not in the API.
| Where | Muse Wonder | Muse Wonder Full | Price |
|---|---|---|---|
| Meta Model API | Yes, as muse-wonder-1.0 (preview) | No | $4 input and $20 output per 1M tokens |
| Muse Code | Every plan | Power and Maximum plans | $5, $15, $50 or $200 a month |
| Muse app | Announced as coming soon, subscribers first | Not announced | Free tier; $20 or $100 a month |
| OpenRouter | Yes, as meta/muse-wonder | No | Set by OpenRouter |
| Meta AI app, WhatsApp, Instagram | No; Meta AI runs Muse Spark | No | Free |
What Muse Wonder is good at, and where it trails
Active Memory
Meta's MemoryBench puts Muse Wonder at 85.6, ahead of Claude Opus 5.5 (81.5) and GPT-6 Astra (81.4). It is Meta's own test, so treat it as a claim.
Active Focus on long tasks
Muse Wonder Full leads Meta's table on long coding work: 79.0 on DeepSWE 1.1 and 57.5 on SWE-Marathon, against 74.2 and 46.3 for Claude Opus 5.5.
Second in Text Arena
On October 13 Muse Wonder (max) entered Arena's text leaderboard at second place with 1513, 12 points behind Gemini 4 Argon. Muse Wonder Full reached sixth in WebDev with 1726.
1M-token context, same price as Opus
$4 and $20 per million input and output tokens, the same list price as Claude Opus 5.5, with no surcharge for long context.
Where Claude Opus 5.5 leads
Opus scores higher on the Artificial Analysis index (58 against 52), on ProgramBench (70.5 against 50.0) and on Terminal-Bench (59.6 against 52.0).
What early users criticise
The best scores belong to Muse Wonder Full, which the API does not offer. Hacker News users found taste and judgment behind Opus, and factual recall is weak.
Muse Wonder benchmarks vs Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1
From Meta's model card, at max reasoning effort (SWE-Marathon for Muse Wonder Full at xhigh). Personal Superintelligence Index and MemoryBench are Meta's own tests. The Artificial Analysis index is the independent v4.3 score; API prices are per 1M input and output tokens.
| Muse Wonder | Muse Wonder Full | Claude Opus 5.5 | GPT-6 Astra | Claude Fable 5.1 | |
|---|---|---|---|---|---|
| Personal Superintelligence Index | 54.3 | 59.1 | 55.2 | 54.5 | 49.9 |
| MemoryBench | 85.6 | 85.5 | 81.5 | 81.4 | 79.0 |
| DeepSWE 1.1 | 78.8 | 79.0 | 74.2 | 74.1 | 67.4 |
| SWE-Marathon | 53.1 | 57.5 | 46.3 | 42.5 | 45.6 |
| Terminal-Bench 4.0-AA | 52.0 | 57.1 | 59.6 | 59.1 | 52.0 |
| ProgramBench | 50.0 | 59.0 | 70.5 | 52.5 | 55.5 |
| OSWorld V2.1, partial credit | 72.6 | 78.4 | 81.8 | 75.8 | 80.7 |
| HealthBench Pro | 60.5 | 62.3 | 65.6 | 65.3 | 62.1 |
| Vals Finance Agent V2 | 56.6 | 59.1 | 58.6 | 53.5 | 58.9 |
| Artificial Analysis index | 52 | 56 | 58 | 53 | 53 |
| API price | $4 / $20 | Subscription only | $4 / $20 | Not confirmed | $10 / $50 |
Official Muse Wonder sources and independent tests
Meta's own pages first, then independent rankings and coverage, all read on October 20, 2026.
How to use Muse Wonder
- 1
Through Meta's API
Create a key at dev.meta.ai and call muse-wonder-1.0 with the OpenAI or Anthropic SDK, or test prompts in the playground at dev.meta.ai/playground. The quickstart is above.
- 2
In Muse Code
Install Meta's coding agent and pick a plan: Everyday ($5) and High ($15) include Muse Wonder; Power ($50) and Maximum ($200) add Muse Wonder Full.
- 3
In the Muse app
Meta says Muse Wonder is coming to its Muse app soon, for subscribers first. It is not in the Meta AI app, WhatsApp or Instagram, which run Muse Spark.
Prompts to try with Muse Wonder
Muse Wonder is built for long, multi-step work. The last two turn its answer into an image prompt you can paste into Nano AI.
Long task with memory
Remember that I'm planning a small bakery in Lisbon with a budget of 40,000 euros. Draft a 12-week launch plan, list what you still need to know from me, and keep track of my answers in later messages.
Code from a screenshot
Here is a screenshot of a pricing page. Rebuild it as a single responsive HTML file with Tailwind CSS, then list anything you had to guess.
Read a long PDF
Read the attached annual report and give me the five numbers that changed most since last year, each with its page number.
Write image prompts
Write three detailed image prompts for a product photo of a matte black water bottle on a mossy rock beside a waterfall: one realistic, one minimalist, one isometric. Describe the light, lens and background, and keep each under 60 words.
Plan a poster, then make it
Plan a poster for a Sunday farmers market: give me the headline, two short lines of small text, and a one-paragraph image prompt that includes the exact text in quotation marks.
Muse Wonder limits and gotchas
What the documentation and early users point out.
- HTTP 400 when turning reasoning off
- Reasoning cannot be disabled. Use minimal effort for the fastest, cheapest answers; reasoning tokens are billed at the output price.
- max effort is not Muse Wonder Full
- Full is a separate model, only in Muse Code's Power and Maximum plans. The benchmark leader in Meta's table is usually Full.
- The OpenAI SDK says the key is missing
- It does not read MODEL_API_KEY. Pass api_key and base_url when you create the client.
- Not available in South Korea
- Meta lists neither Muse Wonder model in South Korea, in the API or in Muse Code.
- Memory with zero data retention
- Zero-data-retention setups do not keep server-side conversation state, which limits what Active Memory can carry over.
- Muse Code trains on your work by default
- Subscriptions start in Contributor mode, which lets Meta train on your content. Switch it off in the settings if you need to.
- It will not make images or video
- Output is text only. Use Muse Image in the Meta AI app, or an image model such as Nano Banana 2.1 on Nano AI.
- Weak factual recall
- Artificial Analysis rates its factual knowledge low (25 on AA-Omniscience), so check names, dates and figures.
Muse Wonder can't make pictures
Need images? Try Nano Banana 2.1 on Nano AI
Muse Wonder reads images but only answers in text, and Meta's video model is not out yet. To turn its ideas into pictures, ask it for an image prompt and paste it here: Nano Banana 2.1 makes the image in your browser for 8 credits, with no API key, and GPT Image 2.5, Seedream 5.0 and other image models are in the same menu.
FAQ
Muse Wonder FAQ
Answers to what people search most about Meta's Muse Wonder.
Is Muse Wonder free?
No free API tier has been announced; the API is pay as you go. The cheapest subscription that includes Muse Wonder is Muse Code Everyday at $5 a month. Free access in the Muse app has not been announced.
How much does the Muse Wonder API cost?
$4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Reasoning tokens count as output, there is no long-context surcharge, and web search grounding costs $2.50 per 1,000 queries.
What is the Muse Wonder API model name?
muse-wonder-1.0, at the base URL https://api.meta.ai/v1. It works with the OpenAI SDK (Responses and Chat Completions) and the Anthropic SDK (Messages). On OpenRouter it is meta/muse-wonder.
Can Muse Wonder generate images?
No. It reads images, video and audio but only outputs text. Meta's image model is Muse Image, free in the Meta AI app and $0.01 per image through the API. On Nano AI you can make images with Nano Banana 2.1, GPT Image 2.5 and other models without an API key.
Muse Wonder vs Claude Opus 5.5: which is better?
They cost the same through the API. Opus scores higher on the Artificial Analysis index (58 against 52), ProgramBench and Terminal-Bench; Muse Wonder leads Meta's MemoryBench and DeepSWE, and Muse Wonder Full beats Opus on SWE-Marathon.
What is the difference between Muse Wonder and Muse Wonder Full?
Full is Meta's most capable variant: 56 against 52 on the Artificial Analysis index and ahead on most of Meta's benchmarks. It has no API; you get it with Muse Code's Power ($50) or Maximum ($200) plan.
Is Muse Wonder open source?
No, neither Muse Wonder model has public weights. Meta's open-weights model in the same family is Muse Spark Open.
Is Muse Wonder in the Meta AI app or WhatsApp?
Not as of October 20, 2026. Meta AI, WhatsApp and Instagram run Muse Spark. Meta says Muse Wonder is coming to its separate Muse app soon, for subscribers first.
What are Active Memory and Active Focus?
Active Memory carries important context about you across conversations and retrieves what a task needs. Active Focus keeps the model working on a hard goal, trying other approaches, working through blockers and checking in before important actions.
What is Muse Wonder's context window?
1,048,576 tokens, about one million, with no extra charge for long context.
Is Nano AI affiliated with Meta?
No. Nano AI is an independent image and video generator. It does not offer Muse Wonder; this page links to Meta's official sources.
More from Nano AI
Related pages
Nano AI is an independent service and is not affiliated with, endorsed by or sponsored by Meta. Muse, Muse Wonder and Meta AI are Meta's trademarks. Benchmark scores, prices and availability come from Meta's model card, research blog, newsroom and developer documentation as read on October 20, 2026; independent results are from Artificial Analysis and Arena on the dates stated.