Muse Wonder and Muse Wonder Full are two sizes of the same Meta model family, released together on October 12, 2026. Wonder is the one you can call: muse-wonder-1.0 on the Meta Model API at $4 / $20 per million tokens, and in every Muse Code plan. Wonder Full is the one Meta’s headline benchmarks describe: four points higher on the Artificial Analysis index, ahead on long coding runs and legal work, and available only to Muse Code subscribers paying $50 or $200 a month, with no API and no per-token price.
That split is the whole story for a developer. This post lays out what differs, what does not, what each costs in practice, and how to decide which one your work needs. Start with the Muse Wonder explainer if the family is new to you.
Side by side
| Muse Wonder | Muse Wonder Full | |
|---|---|---|
| Model ID | muse-wonder-1.0 |
muse-wonder-full-1.0 (Muse Code only) |
| API access | Yes: Responses, Chat Completions, Messages | No |
| Muse Code plans | Everyday ($5) and up | Power ($50), Maximum ($200) |
| Per-token price | $4 in / $20 out / $0.20 cached | None published |
| Context window | 1M tokens | 1M tokens |
| Inputs | Text, image, video, audio | Same |
| Artificial Analysis Intelligence Index | 52 | 56 |
| Coding Agents Index | 63 | 68 |
| Hallucination rate (AA) | 27% | 23% |
| Third-party gateways | Vercel, Cloudflare | None |
| Weights | Closed | Closed |
Everything architectural that Meta describes, Active Memory, Active Focus, multimodal input, computer use, sub-agent delegation, applies to both. The difference is scale: Meta says Full performs better on unfamiliar tasks than the Muse Spark trend line predicted, and attributes that to a larger base model.
Where Full pulls ahead
Meta’s model card, run at maximum reasoning effort:
| Benchmark | Wonder Full | Wonder | Gap |
|---|---|---|---|
| Harvey LAB-AA (legal) | 13.1 | 10.0 | +3.1 |
| AutomationBench-AA | 45.4 | 39.1 | +6.3 |
| JobBench | 75.3 | 70.5 | +4.8 |
| SWE-Marathon | 57.5 | 53.1 | +4.4 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | +5.1 |
| HealthBench Pro | 62.3 | 60.5 | +1.8 |
| DeepSWE 1.1 | 79.0 | 78.8 | +0.2 |
The independent run from Artificial Analysis agrees on direction. Full scores 56 to Wonder’s 52 on the composite, 57% to 52% on Terminal-Bench 4.0, 68 to 63 on the Coding Agents Index, and hallucinates less (23% vs 27%). On Meta’s internal FocusBench, the persistence test behind Active Focus, Full leads every model at 54.2% while Wonder sits at 44.9%, behind GPT-6 Astra and Claude Opus 5.5.
One exception: on Meta’s MemoryBench, Wonder edges Full, 85.6 to 85.5. Memory is a wash.
So the gap is real but uneven. On short software tasks (DeepSWE) it vanishes. On long, blocker-heavy work, the category Meta built the family for, it is four to six points. In the Coding Agents Index, Full ties Claude Code running Sonnet 5.5 at 68 and beats Opus 5.5 at 66; Wonder at 63 trails all three.
What each one costs
Muse Wonder is priced per token. Artificial Analysis measured $4.52 per Intelligence Index task at max effort, below Claude Opus 5.5 ($5.98) and above GPT-6 Astra ($3.26). A typical agentic run that resends a growing context will spend most of its budget on cached input at $0.20; the pricing post works the examples.
Muse Wonder Full has no token price, so there is no per-task figure. Its cost is a subscription: $50 a month for Power (20x the Everyday plan’s usage) or $200 for Maximum (80x). Both share five-hour and weekly usage resets. For an individual who codes all day, $50 a month is cheap relative to API spend on any frontier model. For a team that wants Full behind a product, there is no path at any price.
Note the data terms. Muse Code subscriptions default to Contributor mode, where Meta may train on your prompts, code and outputs. API traffic in Standard mode is not used for training. Opting out of Contributor mode is one toggle, but it is on until you flip it.
Which one you can build on
This is the practical line. Anything you ship to users runs on Wonder, because Full has no API, no gateway, and no SDK. The Full numbers in Meta’s launch material are real, and they are also a description of a tool you can use yourself but not resell.
| You want to | Use |
|---|---|
| Call the model from your product | Wonder, via the API |
| Run Claude Code or Codex on a Meta model | Wonder, via the Messages or Responses endpoint |
| Code in a terminal with the strongest Meta model | Full, via Muse Code Power or Maximum |
| Evaluate Meta’s frontier claims | Full, in Muse Code; then test Wonder for what you will deploy |
| Pay only for what you use | Wonder |
Meta’s developer blog says Full is available only through Muse Code subscriptions and gives no timeline for an API. Until that changes, treat benchmark comparisons of “Muse Wonder” against other APIs with care: if the number is Full’s, it does not describe the API model.
What the wall costs you in practice
Three situations where the Wonder-or-Full distinction changes a decision.
You run Claude Code or Codex on Meta’s API. Meta’s coding-agents guide shows both harnesses pointed at api.meta.ai with muse-wonder-1.0 as the main model and Muse Spark 1.3 for subagents and background tasks. That setup gets Wonder, which scores 63 on the Coding Agents Index. The 68 that ties Claude Code on Sonnet 5.5 belongs to Full inside Muse Code. If the benchmark is why you are switching harnesses, you will not get it this way.
You are comparing vendors for a product. Every Full number in Meta’s launch table is irrelevant to a procurement decision, because no product can call Full. Compare Wonder’s row (52 on the index, 53.1 on SWE-Marathon, 25 on knowledge) with the models you can deploy.
You are not sure which one you are using. In Muse Code, the /models menu shows the active model. A response that describes itself as Wonder is not evidence either way, and the --reasoning-effort flag never changes the model. On the API, the model field in every response says muse-wonder-1.0; there is no Full to confuse it with.
Meta’s own data makes the case for checking: on short tasks the two are within a fifth of a point, so you may not notice which one answered until the task gets long.
Testing the gap on your own work
If you have a Muse Code Power plan, you can measure the gap yourself. Pick one long task, a 30-step refactor or a multi-file review, run it in Muse Code with --model muse-wonder-1.0 and again after switching to Full in /models, and compare the finished result and the number of interventions. That tells you whether the four points matter for your work.
For the API side, Apidog is where you check that Wonder, not a benchmark, meets your bar. Set up the Responses request with the key as a secret variable, run your real prompts at medium and max effort, assert on the output shape, and read usage for the cost. The API guide has the requests. Download Apidog and keep the results next to your other model evals.
Frequently asked questions
Can I get Muse Wonder Full through the API? No. It has no callable model ID, no per-token price, and no third-party gateway. Meta has not announced a date.
Does setting effort to max give me Full? No. Effort controls how long Wonder reasons. The model ID does not change.
Is Full worth $50 a month over Wonder at $5? For solo coding on long tasks, the Coding Agents Index gap (68 vs 63) and the SWE-Marathon gap (57.5 vs 53.1) are meaningful. For short tasks, DeepSWE says they are the same model. For anything you deploy, the question is moot.
Which one does the Muse app use? Meta says Muse Wonder will power app experiences for subscribers first. It has not said the app gets Full.
Where this leaves you
Build on Wonder, because it is the only one you can build on. Use Full in Muse Code if you code all day and want Meta’s strongest model under your hands. And when a chart says Meta beat Opus 5.5 on a coding benchmark, check which Wonder it means before you plan around it.
