Meta released Muse Wonder Full on 12 October 2026 as the top of its 2-model Wonder line, tested at max reasoning effort and available only inside Muse Code. It suits developers on a Power or Maximum plan who want Meta's best long-coding scores, and it replaces nothing on the API side, where Muse Wonder remains the only option.
What changed
- · Release Meta releases Muse Wonder Full for Muse Code subscribers (muse-wonder-full-1.0) source
Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index v4.3.2 and is the higher-capability variant in Meta's Muse Wonder line. It is available only to Power Usage and Maximum Usage subscribers in Muse Code, takes text, image, video and audio input, and returns text.
Provider: Meta AI · Family: Muse
Context window: 1,000,000 tokens
Input modalities: text, image, video, audio, pdf, tool-calls · Output: text, tool-calls
About Muse Wonder Full
Muse Wonder Full is the higher-capability variant of Muse Wonder, released by Meta AI on 12 October 2026. Meta describes the pair as its most capable models, trained at its Prometheus data center on the groundwork of Muse Spark (see all Meta models). Unlike Wonder, Full is not on the Meta Model API: it runs only in Muse Code, Meta's terminal coding agent, for subscribers on the Power Usage or Maximum Usage plans. Meta has not published a parameter count, architecture type or training cutoff.
Meta's comparison table, run at max reasoning effort, has Full ahead of every model listed on DeepSWE v1.1 at 79.0, with the figures taken from a Datacurve run, against 74.2 for Claude Opus 5.5 and 74.1 for GPT-6 Astra. On SWE-Marathon it scores 57.5 at xhigh effort, but that row mixes coding tools (OpenAI Codex and Claude Code for the rivals), so it is not a controlled comparison. It does not win everywhere: Opus 5.5 is ahead on Terminal-Bench 4.0, ProgramBench and the partial-credit OSWorld V2.1 score, and Artificial Analysis independently puts Full at 56 on its Intelligence Index v4.3.2, below the 58 that Meta's table lists for Opus 5.5.
The model card's risk section is written about Full, the most capable variant, and Meta applies its conclusions to the whole family. Without mitigations Meta does not rule out reaching its high-risk threshold in chemical and biological misuse, so it applies safety post-training and system-level classifiers; cybersecurity and loss of control rate moderate or lower even before mitigations. Reported results include 88.4% on CyberGym, 58.9% on ExploitBench, 98.7% on the BioTIER English refusal test and a 0% harmful-action rate on Agentic Misalignment. On ExploitGym it captured 92 flags, but only 21 used the intended vulnerability, and that smaller figure is the one comparable with the published ranking. Meta says a fuller safety report is coming.
Full takes text, image, video and audio input per the model card and returns text, with a 1M-token window per Meta's model page, and Muse Code supports max reasoning effort for it. Muse Code subscriptions default to Contributor mode, which gives more usage in exchange for Meta using your content to train future models; you can turn it off in settings. Artificial Analysis lists Full with no speed or cost-per-task data because there is no public API to benchmark. Meta's announcement names Muse Wonder, not Full, as the model that will soon power experiences in the Muse agent app.
Full is for developers who already pay for Muse Code at the top tiers and want Meta's best coding and agent scores. It is not an option for teams that need API access, cloud-marketplace deployment or predictable per-token billing, and Opus 5.5 still leads on several tests. For API work use Muse Wonder; for cross-vendor choices weigh Claude Fable 5.1 and, for a lower token price, DeepSeek-V4-Pro or Qwen3.8-Max, then see the guide to choosing an LLM and the AI coding assistants roundup.
Pricing
Meta does not sell Muse Wonder Full by the token, and its pricing table shows no per-token rate for it. It is included in two Muse Code plans: Power Usage at $50.00 per month, which Meta lists as 20x the usage of the Everyday plan, and Maximum Usage at $200.00 per month, listed as 80x. Plans renew monthly, upgrades prorate, and a cancelled plan is not refunded unless the law requires it. Full is not offered on pay-as-you-go; for a per-token price see [Muse Wonder](/hub/models/muse-wonder).
| Tier | Rate |
|---|---|
| Power Usage (Muse Code) | $50.00 per month. Includes Muse Wonder Full, early access to new features and higher file uploads. |
| Maximum Usage (Muse Code) | $200.00 per month. Everything in Power Usage with a larger usage allowance. |
Key Features
- Subscription-only access: Included in the Power Usage and Maximum Usage plans of Muse Code; the Everyday and High plans and the Meta Model API do not offer it.
- Max reasoning effort: Muse Code supports the max level for Full, the setting Meta used for the results in its comparison table.
- Multimodal input: Accepts text, image, video and audio with text-only output, per the model card.
- Long-horizon coding: Meta reports the top DeepSWE and SWE-Marathon results in independent third-party runs.
- Active Memory and Active Focus: Carries context across conversations and keeps working through blockers, asking for clarification and checking in before important actions, per Meta.
Pros
- Top DeepSWE row in Meta's table, taken from a Datacurve run rather than a Meta-run score.
- Flat monthly pricing keeps heavy agent sessions from metering by the token.
- Publishes cyber, chemical and biological, and loss-of-control results in its model card, with the ExploitGym caveat stated.
Cons
- Not available on the API, on pay-as-you-go or on any cloud marketplace.
- Trails Claude Opus 5.5 on terminal tasks, ProgramBench and the Artificial Analysis index in Meta's own table.
- Contributor mode is on by default on subscriptions, so Meta may train on your code until you switch it off.
- Artificial Analysis recorded no speed or cost-per-task figure for it.
Benchmarks
- AA Intelligence Index: 56 cited: Artificial Analysis · 14 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does Muse Wonder Full actually cost?
Meta does not sell it by the token. You get it inside Muse Code's Power Usage plan at $50.00 per month or the Maximum Usage plan at $200.00 per month, which Meta lists as 20x and 80x the Everyday plan's usage. Plans renew automatically and cancelled plans are not refunded unless the law requires it. For a per-token price you would use Muse Wonder instead.
Does Muse Wonder Full beat Claude Opus 5.5?
Not across the board. In Meta's table Full leads on DeepSWE and the third-party SWE-Marathon run, while Opus 5.5 scores higher on Terminal-Bench 4.0, ProgramBench and computer-use partial credit, and the Artificial Analysis index puts Full below the figure Meta lists for Opus 5.5. Choose Full for long repository changes inside Muse Code and Opus 5.5 where terminal work or from-scratch programs dominate.
Is Muse Wonder Full open source?
No. It is a proprietary model that runs only inside Muse Code, and Meta's docs list no weights for it or for Muse Wonder. Meta's open-weight model is [Muse Glimmer](/hub/models/muse-glimmer), which you can host yourself. Use is covered by Meta's Model API terms, its acceptable use policy and the Muse Supplemental Terms of Service, and Meta says the models are not intended for people under 18.
Does Muse Wonder Full train on your code?
By default, yes: Muse Code subscriptions start in Contributor mode, which gives you more usage in exchange for letting Meta use your prompts, code, files and responses to train future models. You can switch it off at any time, after which future content is not used. Standard mode keeps content out of training, and Meta's pages do not state a retention period.
Who is Muse Wonder Full for, and who should skip it?
It suits developers already working in Muse Code who hit the limits of lower plans and want Meta's top scores on long coding tasks. Skip it if you need an API, cloud-marketplace deployment or per-token billing, where Muse Wonder or a rival is the route, or if the Contributor-mode default conflicts with client confidentiality terms.
Top Alternatives
- Muse Wonder: Pick Muse Wonder for pay-per-token API access; pick Muse Wonder Full if you hold a Power or Maximum Muse Code plan and want Meta’s top scores.
- Claude Opus 5.5: Pick Muse Wonder Full for long repository coding in Muse Code; pick Claude Opus 5.5 where Meta's own table has it ahead on terminal and from-scratch tasks.
- GPT-6 Astra: Pick Muse Wonder Full if you already live in Muse Code; pick GPT-6 Astra if you need API access or Codex tooling.
- Claude Fable 5.1: Pick Muse Wonder Full for the DeepSWE lead in Meta's table; pick Claude Fable 5.1 if the computer-use rows matter more.