Muse Wonder
Muse Wonder is the model Meta calls frontier-class, released on Oct 12, 2026, sold per token on the Meta Model API (still in preview) and included in Muse Code. It costs US$ 4 per million input tokens and US$ 20 per million output, exactly the price of Claude Opus 5.5 and 40% of what Claude Fable 5.1 and GPT-6 Astra charge for each. In Meta's own table it beats Opus 5.5 on 5 of 14 rows, with its edge in long coding work, memory and legal work. The stronger version, Muse Wonder Full, is not on the API: it only comes with a Muse Code subscription.
Muse Wonder in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Meta (Meta Superintelligence Labs)
- API identifier
muse-wonder-1.0, in preview- Full version
muse-wonder-full-1.0, only with a Muse Code subscription (the US$ 50 and US$ 200 a month plans); not on the API- Release
- Oct 12, 2026
- Input
- US$ 4 per million tokens
- Output
- US$ 20 per million tokens
- Cache read
- US$ 0.20 per million
- Long context
- no surcharge: the same price with an empty or a full window
- Web search
- US$ 2.50 per thousand searches, on top of tokens
- Contributor mode
- not available for Muse Wonder on pay-as-you-go; Standard only, which does not train on your content
- Context window
- 1,048,576 tokens
- Max output
- not published
- Knowledge
- cutoff not published
- Reasoning
- effort from minimal to max; max does not switch to the Full version
- Input and output
- text, image, video, audio and PDF in; text out
- Weights
- not published; Meta does not disclose size or parameter count either
- Where it runs
- Meta Model API and Muse Code (every plan); the Muse agent will get it first for subscribers
- Audience
- Meta says the model is not intended for people under 18
Where Muse Wonder does well, where it falls behind and who it is for
Where it does well
- Long coding work: 78.8 on DeepSWE 1.1 and 53.1 on SWE-Marathon, against 74.2 and 46.3 for Claude Opus 5.5, in Meta's figures. It also beats GPT-6 Astra and Fable 5.1 on both.
- Memory: 85.6 on MemoryBench, the top score in the table, against 81.5 for Opus 5.5, on an internal Meta test.
- Legal work: 10.0 on Harvey LAB-AA, against 4.2 for Opus 5.5 and 8.6 for GPT-6 Astra.
- Price: the same as Opus 5.5 and 40% of Fable 5.1 and GPT-6 Astra (US$ 10 and US$ 50), with no surcharge anywhere in the 1,048,576 token window.
- Data on the API: pay-as-you-go only offers Standard mode, where Meta does not train on your content.
Where it falls behind
- ProgramBench: 50.0 against 70.5 for Opus 5.5, the lowest score in the row and 20.5 points behind.
- Computer use: 72.6 on OSWorld 2.1 (partial, the primary metric), against 81.8 for Opus 5.5. Also the lowest in the row.
- Terminal and health: 52.0 against 59.6 for Opus 5.5 on Terminal-Bench 4.0-AA, and 60.5 against 65.6 on HealthBench Pro.
- Same price, fewer wins: it trails Opus 5.5 on 9 of the 14 rows in Meta's own table.
- Gaps: it is in preview, with no published knowledge cutoff or max output, and Meta lists visual and spatial reasoning, especially 3D, as a limitation.
Who it makes sense for
- Coding agents that work for hours in a large repository, where Meta shows it ahead on DeepSWE and SWE-Marathon.
- Assistants that must remember a long conversation full of distractions, which MemoryBench measures at 64 thousand to 100 thousand tokens.
- Multi-step legal tasks, with a human review of every deliverable.
- Teams paying for Fable 5.1 or GPT-6 Astra that want to test a model at 40% of the per-token price.
- Muse Code subscribers: Wonder is in every plan, from US$ 5 a month.
When to pick another model
- Agents that operate the screen or the terminal: Opus 5.5, at the same price, scores 81.8 against 72.6 on OSWorld 2.1 and 59.6 against 52.0 on Terminal-Bench.
- High volumes of simple tasks: Muse Spark 1.3 costs US$ 1.25 and US$ 4.25, and Wonder is 3.2 times that on input and about 4.7 times on output.
- Health questions: 60.5 on HealthBench Pro, the lowest score among the five models in the table.
- Visual and spatial reasoning, such as 3D tasks, which Meta lists as a limitation in the model card.
- Confidential code in Muse Code on default settings: Contributor mode is on and lets Meta train on your content.
What Meta published at launch
Meta released Muse Wonder on Oct 12, 2026 as a much larger and more advanced family than Muse Spark, in two models: Muse Wonder, for fast frontier performance, and Muse Wonder Full, the full version. The announcement highlights Active Memory, which carries context across conversations, and Active Focus, which keeps the model on task and has it confirm before an important action.
Meta's table has 14 rows, with every model at max effort. 13 are shown here: the last one is the Artificial Analysis index, an independent leaderboard the AI Benchmark does not reproduce. Against Opus 5.5, which costs the same, Muse Wonder comes out ahead on 5 rows (Harvey LAB-AA, GDP.pdf by 0.2 points, MemoryBench, DeepSWE and SWE-Marathon) and behind on 9, counting that index.
| Muse Wonder max | Muse Wonder Full max | Opus 5.5 max | Fable 5.1 max | GPT-6 Astra max | |
|---|---|---|---|---|---|
| PSI internal index of seven tests |
54.3 | 59.1 | 55.2 | 49.9 | 54.5 |
| Harvey LAB-AA legal work |
10.0 | 13.1 | 4.2 | 6.4 | 8.6 |
| Vals Finance Agent V2 finance |
56.6 | 59.1 | 58.6 | 58.9 | 53.5 |
| HealthBench Pro health, no tools |
60.5 | 62.3 | 65.6 | 62.1 | 65.3 |
| AutomationBench-AA process automation |
39.1 | 45.4 | 41.2 | 32.1 | 41.6 |
| GDP.pdf professional agent, Surge test |
30.8 | 33.2 | 30.6 | 27.6 | 34.2 |
| JobBench office tasks |
70.5 | 75.3 | 75.4 | 73.1 | 66.1 |
| OSWorld 2.1, partial computer use, primary metric |
72.6 | 78.4 | 81.8 | 80.7 | 75.8 |
| MemoryBench memory, internal test |
85.6 | 85.5 | 81.5 | 79.0 | 81.4 |
| Terminal-Bench 4.0-AA terminal coding |
52.0 | 57.1 | 59.6 | 52.0 | 59.1 |
| DeepSWE 1.1 long coding work |
78.8 | 79.0 | 74.2 | 67.4 | 74.1 |
| SWE-Marathon long coding work, no common harness |
53.1 | 57.5 | 46.3 | 45.6 | 42.5 |
| ProgramBench runs that pass 95% of tests |
50.0 | 59.0 | 70.5 | 55.5 | 52.5 |
Table from the Muse Wonder announcement, with every model at max effort except Muse Wonder Full on SWE-Marathon (xhigh). The dot marks the best in each row. The 14th row, the Artificial Analysis index, is left out.
How to read Meta's table
Every figure in the table is published by Meta. Under the methodology in the model card, it uses a third-party or competitor-reported result when one exists; otherwise, the highest of its own run, the self-reported score and the official leaderboard. Refusals score zero. None of these scores is independent: the independent ones are at the top of this page, and as of Oct 13, 2026, Epoch AI had only registered Muse Wonder, with no test measured. Three caveats:
- Internal tests: MemoryBench and FocusBench are Meta's own, and PSI (Personal Superintelligence Index) blends seven tests, two of them internal. On FocusBench, in a separate chart, Wonder scores 44.9%, behind Opus 5.5 (45.8%).
- Best effort: where Meta ran the rivals itself, it warns that prompts, tools and runtime may not be tuned for them.
- SWE-Marathon: GPT-6 Astra ran in Codex and Fable 5.1 in Claude Code, and Meta says the row is not a controlled comparison.
Muse Wonder or Muse Wonder Full
Muse Wonder Full is not on the Meta Model API: it is only in Muse Code, on the US$ 50 and US$ 200 a month plans. Setting effort to max on Muse Wonder does not select Full, which has its own page in the AI Benchmark.
In Meta's table, Full has the top score on 6 of the 14 rows and beats Opus 5.5 on 8. Its gap to Wonder is wide on ProgramBench (59.0 against 50.0) and close to nothing on DeepSWE (79.0 against 78.8); on MemoryBench, Wonder is ahead. The four Muse Code plans run from US$ 5 to US$ 200 a month, all with Wonder; Full is only in the top two.
Muse Wonder, Muse Spark 1.3 or Opus 5.5
Wonder has the Opus 5.5 price on all three lines of the table, and neither charges extra for long context. Against Fable 5.1 and GPT-6 Astra, it costs 40% on input and output, and Astra also rises to US$ 20 and US$ 75 above 272 thousand input tokens.
Muse Spark 1.3 is still on sale at US$ 1.25 and US$ 4.25 and has Contributor mode, which cuts the price to US$ 0.10 and US$ 0.20 in exchange for Meta training on the data; on the API, Wonder does not. Meta left Spark 1.3 out of the Wonder table, and not every shared test compares: on DeepSWE v1.1, 75.4 for Spark (with mini-swe-agent) and 78.8 for Wonder (a Datacurve result).
In practice, Spark 1.3 is still the pick for volume. Between Wonder and Opus 5.5, Meta's table favors Wonder on long coding work, memory and law, and Opus on computer use, the terminal, health and ProgramBench. Before switching, run ten of your team's real tasks on both and compare accuracy and cost per task.
| Muse Wonder Meta | Muse Spark 1.3 Meta | Opus 5.5 Anthropic | Fable 5.1 Anthropic | GPT-6 Astra OpenAI | |
|---|---|---|---|---|---|
| Input | US$ 4 | US$ 1.25 | US$ 4 | US$ 10 | US$ 10 |
| Output | US$ 20 | US$ 4.25 | US$ 20 | US$ 50 | US$ 50 |
| Cache read | US$ 0.20 | US$ 0.15 | US$ 0.20 | US$ 0.25 | US$ 1 |
| Long context | same price | same price | same price | same price | US$ 20 and US$ 75 above 272 thousand |
Meta (Standard mode), Anthropic and OpenAI price tables. For GPT-6 Astra, long context starts above 272 thousand input tokens.
Data, safety and limitations
Pay-as-you-go API access only offers Standard mode, which does not train on your content. In Muse Code, subscriptions come with Contributor mode switched on, which gives more usage in exchange for letting Meta train future models on prompts, code and files. Anyone working on client code should turn it off in settings before the first session.
In the model card, Meta says that without safeguards it does not rule out the model reaching the chemical and biological risk threshold of its safety framework; with safety post-training and classifiers, it rates every domain as moderate risk or lower. It also lists visual and spatial reasoning, especially 3D, as a limitation, recommends human confirmation before irreversible actions and says the model is not intended for people under 18.
Examples of using Muse Wonder
An agent takes on a long coding task in a system with many modules, changes the code and runs the tests until they pass. DeepSWE 1.1 measures this across 113 tasks in 91 repositories: 78.8 for Muse Wonder and 74.2 for Opus 5.5, in Meta's table.
An assistant follows a long conversation, full of detail that does not matter, and has to remember what does. MemoryBench tests this at 64 thousand to 100 thousand tokens with distractors, and Wonder scored 85.6. Since the test is internal to Meta, repeat it with your own conversations.
On Harvey LAB-AA, Muse Wonder meets every criterion on 10.0% of legal tasks, against 4.2% for Opus 5.5. It leads, but 9 in 10 deliverables still miss something, and a lawyer's review remains mandatory.
Where Muse Wonder stands out and where it falls behind, by the numbers
Muse Wonder has published results on 11 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Jan 15, 2026 and Jan 10, 2027 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 11 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| AIME 2024 and 2025 (OTIS) | 99.4% | #14 / 75 | #14 / 168 | 100.0% GPT-5.5 |
| CritPt | 28.3% | #13 / 76 | #13 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 60.3% | #7 / 74 | #7 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 82.1% | #15 / 64 (tied with 1) | #15 / 86 | 93.7% GPT-6 Astra |
| Game puzzles | 51.0% | #12 / 58 | #12 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 84.0% | #11 / 64 | #11 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 70.7% | #13 / 54 (tied with 1) | #13 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 9,432 | #8 / 43 | #8 / 57 | 15,515 GPT-6 Astra |
| Gdp Pdf | 30.8% | #4 / 36 | #4 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 54.2% | #10 / 32 (tied with 1) | #10 / 34 | 83.3% Claude Opus 5.5 |
| Ebr Bench | 31.4% | #12 / 22 | #12 / 27 | 76.2% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
Other models from Meta AI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Muse Spark Open | Oct 15, 2026 | no ECI | n/a | n/a | Open weights (restricted use) |
| Muse Wonder | Oct 12, 2026 | no ECI | 1,540 | n/a | API access |
| Muse Spark 1.3 | Sep 2, 2026 | 156.76 | 1,528 | 81.6 | API access |
| Muse Spark 1.2 | Aug 5, 2026 | 154.90 | 1,509 | 78.0 | API access |
| Muse Spark 1.1 | Jul 9, 2026 | 154.00 | 1,516 | 75.3 | API access |
| Muse Spark | Apr 8, 2026 | 152.13 | 1,508 | n/a | API access |
| Llama 4 Maverick | Apr 6, 2025 | 132.20 | n/a | n/a | Open weights (restricted use) |
| Llama 4 Scout | Apr 5, 2025 | 129.65 | n/a | n/a | Open weights (restricted use) |
| Llama 3.3 70B | Dec 6, 2024 | 127.34 | n/a | n/a | Open weights (restricted use) |
Have you used Muse Wonder? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Muse Wonder
How much does Muse Wonder cost?
On the Meta Model API, US$ 4 per million input tokens, US$ 0.20 for cached input and US$ 20 for output, the same price as Claude Opus 5.5, with no long-context surcharge. Web search costs US$ 2.50 per thousand searches. In Muse Code, it is in every plan, from US$ 5 to US$ 200 a month.
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder (muse-wonder-1.0) is sold per token on the Meta Model API and also runs in Muse Code. Full (muse-wonder-full-1.0) is the higher-capability version and only exists with a Muse Code subscription, on the US$ 50 and US$ 200 a month plans. Setting effort to max on Muse Wonder does not turn on Full.
Is Muse Wonder better than Claude Opus 5.5?
At the same price, it depends on the task. In Meta's table, which includes internal tests and rivals run on a best-effort basis, Opus 5.5 comes out ahead on 9 of the 14 rows and Muse Wonder on 5, including DeepSWE, MemoryBench and Harvey LAB-AA. The independent scores are at the top of this page; as of Oct 13, 2026, Epoch AI had only registered the model, with no test measured.
Does Meta train models on my data?
On pay-as-you-go API access, no: Muse Wonder only comes in Standard mode, where content is not used for training. In Muse Code, subscriptions come with Contributor mode switched on, which lets Meta train on prompts, code and files. You can turn it off in settings.
Can I use Muse Wonder in Brazil?
Brazil is not on the restricted territories list in the Meta Model API geographic policy, updated on Sep 18, 2026, but Meta does not publish a list of supported countries: the API works where Meta AI and billing are available. The Muse agent launched only in the United States.
What is Muse Wonder best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Muse Wonder are: SciCode (#7 of 74); Gdp Pdf (#4 of 36); ProofBench (#11 of 64). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Meta, Muse Wonder announcement (with the benchmark table and the FocusBench chart)
- Meta, Muse Wonder model card (PDF, methodology and safety)
- Meta, Muse Wonder model page (identifiers, pricing and text table)
- Meta, Meta Model API models (documentation)
- Meta, Meta Model API pricing and rate limits (documentation)
- Meta, Muse Code plans
- Mark Zuckerberg, launch post on X (Oct 12, 2026)
- Meta, countries where the Model API works (help center)
- Meta, Meta Model API geographic use policy
- Meta, Muse Spark 1.3 announcement (Spark 1.3 DeepSWE score)
- Meta, Muse Spark 1.3 evaluation methodology (PDF)
- Anthropic, pricing (Opus 5.5 and Fable 5.1)
- OpenAI, pricing (GPT-6 Astra)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking