On October 12, Meta released two models, Muse Wonder and Muse Wonder Full, the first top-tier models from its superintelligence lab. By Meta's own table, the bigger Full beats Opus 5.5 in 8 of 14 tests, but it is sold only through a Muse Code subscription. The base Wonder, which you can use through the API at Opus 5.5's price (on GPTunneL both cost $0.012 and $0.06 per 1K tokens), loses to it in 9 of 14 tests.
Below: what was released, what Meta's tests show and how to read them.
What was released and where it's available
| Parameter | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Where available | Meta Model API and Muse Code | Muse Code subscribers only, not in the API |
| Price | on GPTunneL $0.012 / $0.06 per 1K tokens; ID muse-wonder | Muse Code subscription, $50 or $200 a month (not on GPTunneL) |
| Context | 1M tokens | 1M tokens |
| Input / output | text, images, video, audio / text | same |
On our price list Opus 5.5 costs the same as Wonder, while GPT 6 Astra and Fable 5.1 cost $0.02 and $0.10 per 1K tokens. In the Muse Code subscription, the Power tier ($50) gives 20 times the usage of the base Everyday plan, and Maximum ($200) gives 80 times. The weights are not public, so you can't self-host the model. Meta promises to bring Wonder into the Muse agent later, subscribers first.
What Meta's table shows
Meta compared Wonder with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, all at maximum effort. Below we keep the closest rival by price, Opus 5.5. All figures come from the model card; the best of the three is in bold.
| Test | What it measures | Wonder | Wonder Full | Opus 5.5 |
|---|---|---|---|---|
| Personal Superintelligence Index | Meta's own index, personal and work tasks | 54.3 | 59.1 | 55.2 |
| Harvey LAB-AA | an agent's legal work | 10.0 | 13.1 | 4.2 |
| Vals Finance Agent V2 | financial analyst work | 56.6 | 59.1 | 58.6 |
| HealthBench Pro | open-ended health questions | 60.5 | 62.3 | 65.6 |
| AutomationBench-AA | end-to-end business workflows | 39.1 | 45.4 | 41.2 |
| GDP.pdf | professional work with PDFs | 30.8 | 33.2 | 30.6 |
| JobBench | tool use across professions | 70.5 | 75.3 | 75.4 |
| OSWorld 2.1 | computer use | 72.6 | 78.4 | 81.8 |
| MemoryBench | remembering in long dialogues | 85.6 | 85.5 | 81.5 |
| Terminal-Bench 4.0 | agent in the command line | 52.0 | 57.1 | 59.6 |
| DeepSWE 1.1 | long coding tasks | 78.8 | 79.0 | 74.2 |
| SWE-Marathon | ultra-long projects | 53.1 | 57.5 | 46.3 |
| ProgramBench | writing a program from scratch | 50.0 | 59.0 | 70.5 |
| Artificial Analysis Index | composite score | 52 | 56 | 58 |
The base Wonder, which you can use in the API, is ahead of Opus 5.5 in five rows: legal tasks, PDF work, memory and long coding (DeepSWE and SWE-Marathon). In the other nine, Opus is ahead. The biggest gaps are ProgramBench (70.5 vs 50.0), OSWorld (81.8 vs 72.6) and Terminal-Bench (59.6 vs 52.0). Full is ahead of Opus in eight rows out of 14, but it isn't in the API.
How to read these numbers
- The best numbers belong to the version that isn't in the API. The headline results are Full's. The base Wonder you can plug in is much weaker.
- How scores were picked. Meta says that where there is no comparable third-party result, it takes the best of three: its own run, a self-reported score or an official leaderboard. For other vendors' models, runs are "best effort", and prompts and tools may not be tuned optimally.
- Different conditions in one row. In SWE-Marathon, Astra's and Fable's results were taken in Codex and Claude Code, so it isn't a comparison in a common environment. Full is shown there at xhigh effort.
- Own tests. The Personal Superintelligence Index and MemoryBench are Meta's internal tests. The Artificial Analysis index (52 vs 58) and Terminal-Bench have independent verification.
- Refusals. Meta says a model's refusals are scored as zero rather than dropped.
- The legal test. Full's best result on Harvey LAB-AA is 13.1: a lot next to Opus's 4.2, but low in absolute terms. VentureBeat reads this as a sign that agents still can't reliably produce legal deliverables.
What's new beyond the numbers
Meta highlights two features. Active Memory: the model carries what it knows about the user across conversations and pulls up what's relevant to the task. Active Focus: on a hard goal it tries other approaches, works around blockers, and when needed asks for clarification and checks in before important actions. There are numbers for memory (85.6 vs 81.5 for Opus), but the test is internal. For Active Focus, Meta cites its own FocusBench.
Safety and limitations
Meta calls Wonder its "most aligned model to date" and promises its strongest safeguards yet against misuse in chemistry, biology and cybersecurity. A full report is promised later. The model card has details that the announcement doesn't:
- without mitigations, Meta doesn't rule out that the model could reach a high risk threshold in chemistry and biology, so extra filters are on;
- on the ExploitGym test, Full captured 92 flags, but only 21 through the intended vulnerability; the other 71 came through unintended ones. The ranking counts 21;
- the weak spot is vision and spatial reasoning, especially 3D tasks;
- the model was not evaluated on all languages in its training data, and it is not intended for people under 18.
What this means for you
The per-token price of Wonder and Opus 5.5 is the same, but by Artificial Analysis's estimate, Wonder uses fewer tokens than Opus 5.5 to run its index. So per task Wonder is on average about a quarter cheaper. The choice comes down to quality and cost on your own task.
- Memory about the user, legal tasks, big coding projects. The base Wonder is worth testing: by Meta's data it is ahead of Opus 5.5 here. On financial tasks (Vals), Opus is slightly ahead: 58.6 vs 56.6.
- Computer use, terminal work, programs from scratch, medical questions. Opus 5.5 is ahead by 5-20 points here. For these tasks Wonder is not yet a replacement.
- Full's highest scores. You can use it only inside a Muse Code subscription, not through the API and not in your own services.
There is also a condition for business: according to VentureBeat, Muse Code subscriptions default to Contributor mode, in which Meta may use submitted data to improve its models, and zero data retention is not on by default.
Meta's figures are self-reported. Before switching, run your own set of requests on both models.
Where to try
Muse Wonder is now in the GPTunneL catalog. This is the base API version: Full is only available with a Muse Code subscription. The context is 1,048,576 tokens and the price is $0.012 per 1K input tokens and $0.06 per 1K output tokens, exactly the same as its main rival Claude Opus 5.5: a request with 100,000 input tokens and 10,000 output tokens costs $1.80 on either. So the easiest way is to compare them on your own task in one chat; for details on Opus, see the Opus 5.5 breakdown. From the same lineup there is the cheaper Muse Spark 1.3, covered in a separate article. No VPN needed, local payment methods, no subscription, one shared balance. Current prices are on the pricing page.
Sources
- Muse Wonder announcement, Meta, 10/12/2026 - models, availability, Active Memory and Active Focus.
- Muse Wonder Model Card, Meta, 10/12/2026 - test table, methodology, safety, limitations.
- Meta releases its first frontier model, Muse Wonder, TechCrunch, 10/12/2026 - API price and access terms.
- Meta’s new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API, VentureBeat, 10/12/2026 - Artificial Analysis index, plans, data terms.