Meta has a frontier AI model, and this time you cannot download its weights. Meta Muse Wonder arrived on October 12, 2026, from Meta Superintelligence Labs, built for coding, memory and agent tasks that take more than a quick reply. You can use Wonder through the Meta Model API today at $4 per million input tokens and $20 per million output tokens.
Want the stronger Wonder Full? You need a paid Muse Code subscription. If you use the regular Muse app, access is coming later, with subscribers getting it first.
You see, the model you can call through an API and the model behind Meta’s highest scores are different. Here are the details you need before choosing.
| Item | Detail |
|---|---|
| Released | October 12, 2026 |
| Made by | Meta Superintelligence Labs |
| Models | Muse Wonder and Muse Wonder Full |
| Context window | 1,048,576 tokens for Wonder |
| Inputs/outputs | Text, images, audio, video and PDFs in. Text out |
| API price | $4 input, $20 output and $0.20 cached input per million tokens |
| Wonder Full access | Paid Muse Code subscriptions |
| Consumer access | Muse app rollout coming, subscribers first |
| Open weights | Neither Wonder model |
| Headline features | Cross-conversation memory, persistent task completion and coding |
What Meta Muse Wonder Actually Is
Muse Wonder is a model family built to carry out work over multiple steps. Meta’s examples include dealing with an insurance claim, making sense of a lease and creating an app. To understand what you get, start with the two models and their memory features.
Wonder and Wonder Full are larger than Muse Spark, Meta’s earlier model built for quicker responses. They were trained at the Prometheus data center in central Ohio, and reporting identifies Watermelon as their internal codename.
Active Memory
You should not have to explain the same important details every time you return. Meta says Active Memory lets the models retain useful information between conversations.
The aim is continuity over days, weeks and months. That is Meta’s claim, though, rather than a guarantee that every detail you provide will be remembered correctly.
Sticking With Hard Tasks
What happens when a task goes wrong halfway through? Meta says its models can try another method, recover after mistakes and ask you for missing information.
They are also designed to check with you before important actions. The intended benefit is an agent that keeps working through a difficult assignment instead of stopping at the first failed step.
How to Get Access Today
Your route depends on whether you want an API, a coding subscription or the everyday Muse app. Wonder is available to developers now, while the consumer rollout is still ahead. Choose the route that matches your work.
For developers using the API
- Use the Meta Model API and select muse-wonder-1.0.
- Connect through compatible developer tools. The API supports the OpenAI SDK.
- Alternatively, use Wonder through OpenRouter, where Meta supplies the model.
For coders who want Wonder Full
- Choose a paid Muse Code plan that includes Wonder Full.
- Power Usage costs $50 monthly and provides 20 times the Everyday plan allowance.
- Maximum Usage costs $200 monthly and provides 80 times that allowance.
Wonder Full has no published pay-as-you-go price and is unavailable through the public API.
For everyday Muse app users
- Wait for Meta’s announced Wonder rollout, which starts with subscribers.
- Check which Muse subscription you have. Reported app plans are Free, Power at $20 monthly and Maximum at $100 monthly.
- Check eligibility. The existing Muse agent launched for US users aged 18 and over.
Keep the two subscription products separate. Muse app prices differ from Muse Code prices, and having the app today does not mean Wonder has arrived in it.
Muse Wonder Pricing Compared With Rivals
Wonder costs the same per token as Claude Opus 5.5, while GPT-6 Astra and Claude Fable 5.1 charge more. But your final bill also depends on how many tokens a task consumes. Start with the published rates.
| Model | Input per 1M | Output per 1M | Cached input |
|---|---|---|---|
| Muse Wonder | $4 | $20 | $0.20 per 1M |
| Muse Spark 1.3 | $1.25 | $4.25 | Not listed |
| Claude Opus 5.5 | $4 | $20 | Not listed |
| GPT-6 Astra | $10 | $50 | Not listed |
| Claude Fable 5.1 | $10 | $50 | Not listed |
Artificial Analysis reported average Index task costs of $4.52 for Wonder, $5.98 for Opus 5.5 and $3.26 for Astra. So Astra’s higher token price did not make it the most expensive on that evaluation. Your own jobs need the same cost check.
Wonder’s rates stay unchanged across context lengths. Web search adds $2.50 per 1,000 queries, with API limits of 3,000 requests and four million tokens each minute.
How Meta Muse Wonder Scores Against GPT-6, Gemini 4 Argon and Claude
Wonder Full ranks above Wonder on the Artificial Analysis Index, but neither leads every test. You need to look beyond the overall score to see which results matter for your work. Start with the ranking, then compare individual tasks.
| Model | Artificial Analysis Index score |
|---|---|
| Claude Opus 5.5 | 57.6 |
| Muse Wonder Full | 55.8 |
| Claude Fable 5.1 | 53.4 |
| GPT-6 Astra | 52.7 |
| Gemini 4 Argon | 52.6 |
| Muse Wonder | 52.4 |
| GPT-6.1 Sol | 51.8 |
Meta’s comparison at maximum reasoning gives you a more detailed view.
| Benchmark | Wonder Full | Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Harvey LAB-AA | 13.1 | 10.0 | 8.6 | 4.2 |
| AutomationBench-AA | 45.4 | 39.1 | 41.6 | 41.2 |
| JobBench | 75.3 | 70.5 | 66.1 | 75.4 |
| DeepSWE 1.1 | 79.0 | 78.8 | 74.1 | 74.2 |
| SWE-Marathon | 57.5 | 53.1 | 42.5 | 46.3 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 59.1 | 59.6 |
| HealthBench Pro | 62.3 | 60.5 | 65.3 | 65.6 |
- Legal work. Both Wonder models beat Astra and Opus in Harvey LAB-AA.
- Long coding tasks. Both lead those rivals on DeepSWE 1.1 and SWE-Marathon.
- Business workflows. Full leads AutomationBench-AA, while Opus narrowly wins JobBench.
- Terminal and health tasks. Both Wonder models trail Astra and Opus on the listed tests.
- Other weaknesses. Full scores below Fable on OSWorld GUI completion and below Opus on ProgramBench.
These figures mix third-party evaluations with tests run by Meta. Independent replication was unavailable at launch, and Meta’s comparison table provides no Gemini 4 Argon results.
Open or Closed Weights
Neither Wonder model comes with downloadable weights, so you cannot host either on your own hardware. That follows Muse Spark’s closed release in April, after Meta’s earlier Llama releases made weights available. If self-hosting matters to you, the relevant alternative is Muse Glimmer.
Glimmer is a 30-billion-parameter open-weight model released under Apache 2.0 and can run on one GPU. It is a separate model, rather than a downloadable edition of Wonder.
Alexandr Wang says Meta remains committed to open source. Zuckerberg also said in September that an open Muse Spark would arrive soon, but it has not been released. You have no announced date for Wonder weights, so plan around the access available today.
Privacy and Data Use You Should Check First
Before you give Wonder a codebase, a client document or a personal conversation, check how you are accessing it. The API, Muse Code and Muse app have different rules for training data. Here is what you need to know before sending anything sensitive.
- Standard mode. Meta says submissions through this mode are excluded from model training. The pay-as-you-go API operates under this arrangement.
- Contributor mode. Muse Code subscriptions start with this mode enabled, allowing Meta to train using material you submit and responses you receive. That can include your prompts, source code and uploaded files.
- Turning Contributor mode off. Your choice applies to material submitted afterward. It does not retroactively change how earlier submissions are treated, and the terms prohibit confidential or personal material in Contributor services.
- Zero data retention. Approved API organisations can request this option, but exclusions apply. You should check those exclusions before treating it as suitable for your requirements.
- Muse app training. Your app conversations are used for training by default, although you can opt out. Meta separately says those conversations are kept out of its advertising systems.
- The Muse Image warning. Meta withdrew Muse Image three days after release following criticism of public accounts being opted in by default. That was a different product, but it gives you a reason to inspect defaults yourself.
You see, paying for access does not automatically answer the training question. A subscription and an API bill can come with different data terms.
Before your first serious task, take these three steps.
- Check the data mode for the specific service you intend to use.
- Disable Contributor mode before submitting new material if you do not want it used for training, and keep restricted information out of Contributor services.
- Keep human approval in place for consequential agent actions, as Meta recommends.
Who Should Try It and Who Can Wait
Wonder looks worth testing when your work involves long coding assignments or agents handling several connected steps. It is a less certain choice when you need downloadable weights, dependable screen control or a published full safety report. Start with the group closest to your needs.
Worth testing now.
- Developers with an Opus-level budget. Wonder has the same published input and output rates as Opus 5.5. You can compare results and completed-task costs without moving to Astra’s higher per-token rates.
- Legal and finance agent builders. Meta’s published results give you reasons to investigate these uses, including strong legal scores and a finance result where Wonder Full narrowly beats Opus. Test your own assignments before trusting that advantage.
- Coders handling long refactors. The DeepSWE and SWE-Marathon results support trying Wonder on sustained software work. Remember that the stronger Full model requires a Muse Code subscription.
You can wait if these requirements matter more.
- Casual Muse app use. Wonder has not reached the app yet, and subscribers are due to receive it first. You do not need to buy a coding plan just to prepare for that rollout.
- Heavy reliance on screen-based automation. Wonder Full trails Claude Fable 5.1 on the reported GUI completion test. Check reliability on your own browser and app tasks before choosing it.
- Self-hosting or safety documentation requirements. Neither Wonder model has downloadable weights, and Meta’s full safety report is still pending. If either is essential, the launch leaves that requirement unanswered.
The useful question is whether Wonder helps you finish your work. A high benchmark score gives you a reason to test, rather than a reason to skip testing.
What We Still Don’t Know
Meta has announced the models, their main access routes and several performance claims. You still lack details that could affect a purchase or deployment decision. Keep these unanswered points on your checklist.
- Full safety report timing. Meta says a report is coming, but has not provided a publication date. You cannot yet review that promised document.
- Wonder Full API pricing. Full is restricted to Muse Code subscribers, with no published pay-as-you-go rate. You have no confirmed public API offering to budget for.
- Muse app release date and regions. Subscribers are first in line, but Meta has not specified when Wonder will reach the app or where that rollout will begin. The existing agent’s US eligibility does not settle Wonder’s future availability.
- Downloadable weights. Meta’s statements about open source do not include a release schedule for either Wonder model. The earlier promise of an open Muse Spark also remains unfulfilled.
- Independent checks of Meta’s results. Separate replication was unavailable at launch. You therefore cannot assume every published lead will survive testing by others.
- EU availability. Wonder’s EU status is still unresolved. Meta’s refusal to sign the EU AI Act code of practice does not, by itself, tell you when or whether you will get access there.
“Coming soon” can leave plenty of room for uncertainty. Until you have a dated announcement, keep your plans tied to what you can actually use.
What to Watch Next
Your next decision should depend on useful evidence, not the excitement around a launch. Watch for Meta’s full safety report, the Muse app rollout and any changes to access or pricing. Those updates will help you decide whether to test now or return later.
If you build with APIs, keep an eye on pricing from OpenAI, Anthropic and Google too. The task-cost figures already show why token rates alone cannot settle your budget. Compare what you spend to get an acceptable result.
If you need open weights, watch the promised Muse Spark release and any specific announcement about Wonder. Prometheus matters as the training infrastructure behind these larger models, but it does not answer when downloadable versions might arrive.
For now, choose one task you know well and try Wonder against your current model. Check the result, corrections needed and total cost, while keeping the data rules in mind. You do not need to replace your whole workflow at once – one useful test is enough to decide your next step.
FAQs
You pay to use Wonder through the API, and Wonder Full requires a paid Muse Code subscription. The Muse app reportedly offers a free tier, but Meta has not confirmed how Wonder access will work on that tier. Its announced app rollout starts with subscribers.
You cannot download either Wonder model because Meta has kept their weights private. Your available routes rely on Meta’s hosted services. If you need a model you can run yourself, Muse Glimmer is a separate open-weight option released under Apache 2.0.
Wonder Full scores higher than Wonder on the listed Artificial Analysis Index and Meta benchmark comparisons. You can access regular Wonder through the API or Muse Code, whereas Full is limited to Muse Code subscriptions. Meta has not announced a public API price for Full.
You cannot call Wonder the better coding choice across every task. Both Wonder models beat GPT-6 Astra and Opus 5.5 on the listed DeepSWE and SWE-Marathon tests, but trail them on Terminal-Bench. Your own coding assignments are the best next check, especially given the lack of independent replication at launch.
Whether Meta uses your submissions for training depends on the service and data mode you choose. Pay-as-you-go API submissions are excluded, while Muse Code starts in Contributor mode and the Muse app trains on conversations by default. You can opt out, but changing Contributor mode applies only to later submissions.
Meta has not announced an exact date for Wonder’s arrival in the Muse app. It says subscribers will receive access before other users. You should wait for a dated rollout notice rather than assume your current app subscription already includes the model.
