Muse Wonder Full
Muse Wonder Full is the highest-capability version of the Muse Wonder family, released by Meta on Oct 12, 2026, and it runs only inside Muse Code, Meta's terminal coding tool, on the US$ 50 and US$ 200 per month plans. It has no API and no per-token price, so it cannot go into your own product, support channel or integration. In Meta's own table it has the top score in 6 of 14 rows and beats Claude Opus 5.5 in 8, but loses to it on ProgramBench, 59.0 to 70.5. No independent source had measured Full as of Oct 13, 2026.
Muse Wonder Full 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)
- Identifier
muse-wonder-full-1.0, Muse Code subscription only- Release
- Oct 12, 2026
- Where it runs
- only in Muse Code, Meta's terminal coding tool
- Plans that include Full
- Power Usage, US$ 50 per month (20 times the usage of the US$ 5 plan), and Maximum Usage, US$ 200 per month (80 times)
- On the Meta Model API
- not available, no per-token price; max effort on Muse Wonder does not select Full
- Context window
- 1 million tokens at inference, per the model card; the documentation lists 1,048,576 for Muse Wonder
- Input and output
- text, image, video and audio in; text out
- Max output
- not published
- Knowledge cutoff
- not published
- Effort in Meta's tests
- max; xhigh on SWE-Marathon
- Training on your data
- Contributor mode on by default in Muse Code subscriptions; can be turned off in settings
- Intended users
- not intended for anyone under 18, per the model card
- Full safety report
- promised soon; the model card already evaluates Full
Where Muse Wonder Full does well, where it falls behind and who it is for
Where it does well
- Legal work: 13.1 on Harvey LAB-AA, against 8.6 for GPT-6 Astra, the best rival, and 4.2 for Claude Opus 5.5, in Meta's table.
- Long coding work: 79.0 on DeepSWE 1.1, against 74.2 for Opus 5.5, both runs done by Datacurve, outside Meta.
- Automation: 45.4 on AutomationBench-AA, against 41.6 for GPT-6 Astra and 41.2 for Opus 5.5.
- Staying on task: 54.2% on FocusBench, Meta's internal agent test, against 47.7% for GPT-6 Astra and 45.8% for Opus 5.5.
- Above its sibling: ahead of Muse Wonder in every row of the table except MemoryBench (85.5 against 85.6).
Where it falls behind
- No API: no per-token price and no access through the Meta Model API. It runs only in Muse Code, on a US$ 50 or US$ 200 monthly subscription.
- ProgramBench: 59.0 against 70.5 for Claude Opus 5.5, 11.5 points, the widest gap against Full in Meta's table.
- Terminal and computer use: 57.1 on Terminal-Bench 4.0-AA, against 59.6 for Opus 5.5, and 78.4 on OSWorld V2.1, against 81.8 for Opus.
- Training on by default: in subscriptions, Contributor mode, which lets Meta train future models on your content, comes switched on.
- Only Meta has measured it: internal tests, competitors run on a "best-effort" basis and no independent measurement as of Oct 13, 2026.
Who it makes sense for
- Developers who code in the terminal and are willing to pay US$ 50 a month, on Power Usage, for Meta's highest-capability model.
- Long, multi-step coding in a single repository: Full leads DeepSWE 1.1 (79.0) and SWE-Marathon (57.5, at xhigh) in Meta's table.
- Teams already using Muse Wonder in Muse Code that want more quality without switching tools.
- Teams that want to compare before deciding: on Power Usage, Full and Muse Wonder come in the same plan.
When to pick another model
- Your own product, support channel or integration: Full has no API. Within the family, the route is Muse Wonder, at US$ 4 input and US$ 20 output per million tokens.
- Client code or documents with Contributor mode on, the subscription default: turn it off in settings first.
- Tasks like ProgramBench, Terminal-Bench 4.0-AA and OSWorld V2.1, where Claude Opus 5.5 scored higher in Meta's own table.
- Work that depends on visual and spatial reasoning, especially 3D, a limitation the model card states.
- Running the model in the agent or editor your team already uses: Full exists only in Muse Code.
One family, two models: what is different about Full
Meta launched the Muse Wonder family on Oct 12, 2026 with two models. Mark Zuckerberg presented Muse Wonder as the one for fast frontier performance and Full as the full version of the model; the documentation calls it the "higher-capability variant".
The difference that matters most is access. Muse Wonder is on the Meta Model API and in every Muse Code plan; Full is only in the two most expensive Muse Code plans. And setting effort to "max" on the API's Muse Wonder does not select Full, the documentation warns: they are different models.
| Muse Wonder Full muse-wonder-full-1.0 | Muse Wonder muse-wonder-1.0 | |
|---|---|---|
| Where it runs | Muse Code only | Meta Model API and Muse Code |
| Muse Code plans | Power Usage and Maximum Usage | all |
| Price | US$ 50 or US$ 200 monthly subscription | US$ 4 input and US$ 20 output per million tokens |
| Training on your data | Contributor mode on by default in the subscription | pay-as-you-go API: Standard mode only, no training |
| Top score in Meta's table | 6 of 14 rows | 1 of 14 rows |
Meta Model API price on the Standard tier; Muse Wonder cache reads cost US$ 0.20 per million tokens. Muse Code plans from the product page, checked on Oct 13, 2026.
Meta's table: top score in six rows, ahead of Opus 5.5 in eight
Meta compared both versions with Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra, all at max effort. Across 14 rows, Full has the top score in six and Opus 5.5 in another six, including the Artificial Analysis index Meta put in the table (58 against 56). GPT-6 Astra takes GDP.pdf, and Muse Wonder takes MemoryBench.
Against Opus 5.5, Full is ahead in 8 of the 14 rows; the widest gap against it is on ProgramBench, 59.0 against 70.5. Over its sibling, the biggest lead is also on ProgramBench (59.0 against 50.0). On FocusBench, an internal agent test in a separate chart, Full scores 54.2%, the top result, and Muse Wonder 44.9%.
| Muse Wonder Full max | Muse Wonder max | Fable 5.1 max | GPT-6 Astra max | Opus 5.5 max | |
|---|---|---|---|---|---|
| PSI Meta's internal index, seven tests |
59.1 | 54.3 | 49.9 | 54.5 | 55.2 |
| Harvey LAB-AA legal work |
13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Vals Finance Agent V2 finance agent |
59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| HealthBench Pro health, no tools |
62.3 | 60.5 | 62.1 | 65.3 | 65.6 |
| AutomationBench-AA automation |
45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| GDP.pdf agent, sourced from Surge |
33.2 | 30.8 | 27.6 | 34.2 | 30.6 |
| JobBench tasks from 35 office occupations |
75.3 | 70.5 | 73.1 | 66.1 | 75.4 |
| OSWorld V2.1 computer use, partial reward |
78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
| MemoryBench (internal) memory in long conversations |
85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
| Terminal-Bench 4.0-AA terminal work |
57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
| DeepSWE 1.1 113 tasks across 91 repositories |
79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| SWE-Marathon coding, uncontrolled comparison |
57.5 (xhigh) | 53.1 | 45.6 | 42.5 | 46.3 |
| ProgramBench runs passing at least 95% of tests |
59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
| Artificial Analysis Index cited by Meta |
56 | 52 | 53 | 53 | 58 |
Table from the Muse Wonder announcement, identical on Meta's blog and on the model page. All at max effort, except Full on SWE-Marathon, which ran at xhigh. The dot marks the best in each row. On OSWorld V2.1, the row uses the primary metric, partial reward; on the secondary binary metric, Full scores 39.4 and Fable 5.1 leads with 43.1.
How to read Meta's numbers
The model card methodology carries caveats that change how the table reads:
- Internal tests: PSI is Meta's own index, combining seven evaluations, including FocusBench and VBT, both internal too. MemoryBench is internal. None of them can be checked from the outside.
- Competitors on a "best-effort" basis: for the runs it did itself, Meta says prompts, tools and runtime may not be tuned for other companies' models. Refusals score zero.
- SWE-Marathon: Meta says this row is not a controlled comparison. GPT-6 Astra ran in Codex, Fable 5.1 in Claude Code, and Full at xhigh, not max.
- What comes from outside: of the six rows Full leads, only PSI is measured by Meta; the others have outside sources, according to Meta (Artificial Analysis, Vals AI, Datacurve and Abundant AI).
Subscription and data: what you buy in Muse Code
Full has no per-token price: it comes with the top two Muse Code plans, which measure usage as multiples of the entry plan.
In subscriptions, Contributor mode is on by default: Meta describes it as more usage in exchange for training future models on your content. It can be turned off in settings, and for a company that opens client code or documents in the agent, it is the first thing to switch off. On the pay-as-you-go API, Muse Wonder runs only in Standard mode, with no training.
| Price per month | Usage | Muse Wonder Full | |
|---|---|---|---|
| Everyday Usage plan | US$ 5 | 10 to 50 prompts every 5 hours | no |
| High Usage plan | US$ 15 | 5 times Everyday | no |
| Power Usage plan | US$ 50 | 20 times Everyday | yes |
| Maximum Usage plan | US$ 200 | 80 times Everyday | yes |
Meta's Muse Code page, checked on Oct 13, 2026. Muse Wonder is included in every plan.
What the model card says about Full
Meta ran its safety evaluation on Full, the most capable variant, and says the full report is coming soon. What is already published:
- Chemistry and biology: without mitigations, Meta does not rule out the model reaching its framework's risk threshold. With safety post-training and classifiers, every domain sits at "moderate or lower".
- Cyber: on ExploitGym, Full captured 92 flags, only 21 through the vulnerability the test intended. Anyone giving the agent access to a real system has a concrete reason to limit permissions.
- Behavior: 87.3% on MASK and 0% on Agentic Misalignment.
- Stated limitations: visual and spatial reasoning, "especially 3D tasks". Meta recommends human confirmation before irreversible actions.
Examples of using Muse Wonder Full
In Muse Code, the agent gets a large change: understand the code, edit several modules and run the tests until they pass. That is the work DeepSWE 1.1 and SWE-Marathon measure, where Full leads Meta's table.
A Power Usage subscriber has both models. Run ten real tasks on both and compare accuracy and how much of the plan's limit each one uses. Only that test tells you whether Full pays off.
An agency is about to use Muse Code on a client's system. Before the first command, it turns off Contributor mode in settings, because subscriptions ship with it on.
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) |
What people ask about Muse Wonder Full
How can I use Muse Wonder Full?
Only through Muse Code, Meta's terminal coding tool, on the Power Usage (US$ 50 per month) and Maximum Usage (US$ 200 per month) plans. There is no API and no per-token price, and max effort on the API's Muse Wonder does not select Full.
What is the difference between Muse Wonder and Muse Wonder Full?
Full is the family's higher-capability variant and runs only in Muse Code, on the two most expensive plans. Muse Wonder is on the API, at US$ 4 input and US$ 20 output per million tokens, and in every Muse Code plan.
Is Muse Wonder Full better than Claude Opus 5.5?
In Meta's table, Full is ahead of Opus 5.5 in 8 of 14 rows and behind in 6, including ProgramBench (59.0 against 70.5). The figures are Meta's, and Opus 5.5 has an API; Full does not.
Does Meta train models on what I do in Muse Code?
It can. In subscriptions, Contributor mode is on by default and lets Meta use your content to train future models. You can turn it off in settings.
Why does Muse Wonder Full have no independent score?
Because no outside source had measured Full as of Oct 13, 2026: Epoch AI has listed Muse Wonder, but not Full. Every figure for it on this page comes from Meta.
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 family announcement (benchmark table and FocusBench chart)
- Meta, Muse Wonder model card (PDF, methodology and safety)
- Meta, Muse Wonder 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
- Mark Zuckerberg, the two models in the family (thread on X)
- Anthropic, pricing (Claude Opus 5.5)
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