Muse Wonder vs Muse Wonder Full

Muse Wonder and Muse Wonder Full are two sizes of the same Meta model family, released together on October 12, 2026. Wonder is the model developers can call directly: muse-wonder-1.0 is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, and it is included in every Muse Code plan. Wonder Full is the model behind Meta’s headline benchmark results: it scores four points higher on the Artificial Analysis index, leads on long coding runs and legal work, and is available only through Muse Code Power or Maximum subscriptions at $50 or $200 per month. It has no public API and no published per-token price.

Try Apidog today

For developers, that access split is the main decision. This guide compares the two models feature by feature, explains where Full performs better, breaks down the practical costs, and shows how to choose the right model for coding, evaluation, and production use. If the Muse family is new to you, start with the Muse Wonder explainer.

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 above	Power ($50), Maximum ($200)
Per-token price	$4 input / $20 output / $0.20 cached input	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, Artificial Analysis	27%	23%
Third-party gateways	Vercel, Cloudflare	None
Weights	Closed	Closed

The architectural capabilities Meta highlights apply to both models: Active Memory, Active Focus, multimodal input, computer use, and sub-agent delegation. The practical difference is scale. Meta says Full performs better on unfamiliar tasks than the Muse Spark trend line predicted, and attributes that result to its larger base model.

Where Full pulls ahead

Meta’s model-card results below were 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 Artificial Analysis evaluation points in the same direction. Full scores 56 versus Wonder’s 52 on the composite index, 57% versus 52% on Terminal-Bench 4.0, and 68 versus 63 on the Coding Agents Index. It also hallucinates less: 23% versus 27%. On Meta’s internal FocusBench, which measures the persistence behavior behind Active Focus, Full leads every evaluated model at 54.2%, while Wonder scores 44.9% and trails both GPT-6 Astra and Claude Opus 5.5.

Memory is the exception. On Meta’s MemoryBench, Wonder narrowly beats Full, 85.6 to 85.5, which makes that capability effectively a wash.

The performance gap is therefore real but uneven. On short software tasks such as DeepSWE, the difference nearly disappears. On long, blocker-heavy coding work—the category Meta designed the family for—Full leads by roughly four to six points. On the Coding Agents Index, Full ties Claude Code running Sonnet 5.5 at 68 and beats Opus 5.5 at 66, while Wonder trails all three at 63.

What each model costs

Muse Wonder

Wonder uses token-based pricing. Artificial Analysis measured its cost at $4.52 per Intelligence Index task at maximum effort, compared with $5.98 for Claude Opus 5.5 and $3.26 for GPT-6 Astra.

Agentic workflows often resend a growing context, so cached input can dominate the bill. At $0.20 per million cached input tokens, that can substantially reduce the effective cost of long coding sessions. The Muse Wonder pricing guide walks through the cost examples.

Muse Wonder Full

Full has no token price, so there is no meaningful per-task API cost figure. Its price is the subscription:

Plan	Monthly price	Usage allowance
Muse Code Power	$50	20x Everyday
Muse Code Maximum	$200	80x Everyday

Both plans use five-hour and weekly usage resets. For an individual developer who codes throughout the day, $50 per month can be inexpensive compared with API usage on another frontier model. For a team that wants to expose Full behind its own product, however, there is no supported path at any price.

The data terms also matter. Muse Code subscriptions default to Contributor mode, which allows Meta to train on prompts, code, and outputs. API traffic in Standard mode is not used for training. You can turn Contributor mode off with one toggle, but it remains enabled until you do so.

Which model can you build on?

This is the practical dividing line. Anything shipped to users runs on Wonder because Full has no API, no third-party gateway, and no SDK. Full’s launch benchmarks are real, but they describe a tool you can use yourself—not a model you can resell or embed in a product.

Goal	Best choice
Call the model from your product	Wonder, through the API
Run Claude Code or Codex on a Meta model	Wonder, through Messages or Responses
Code in a terminal with Meta’s strongest model	Full, through Muse Code Power or Maximum
Evaluate Meta’s frontier claims	Test Full in Muse Code, then test Wonder for deployment
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 API access. Until that changes, be careful with benchmark tables comparing “Muse Wonder” to other APIs: if the score belongs to Full, it does not describe the API model.

What the access wall costs in practice

You run Claude Code or Codex against 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 handling subagents and background tasks. That configuration gives you Wonder, which scores 63 on the Coding Agents Index.

The 68 score that ties Claude Code running Sonnet 5.5 belongs to Full inside Muse Code. If that benchmark is your reason for changing harnesses, this setup will not deliver it.

You are comparing vendors for a product

Every Full result in Meta’s launch table is irrelevant to procurement because no external product can call Full. For a deployable comparison, use Wonder’s numbers: 52 on the Artificial Analysis index, 53.1 on SWE-Marathon, and 25 on knowledge. Compare those directly against models your product can actually access.

You are unsure which model answered

In Muse Code, open the /models menu to see the active model. A response that calls itself “Wonder” is not reliable evidence of which model produced it, and the --reasoning-effort flag does not switch models.

On the API, the model field in every response identifies muse-wonder-1.0. There is no Full model on the API to confuse it with.

Meta’s own benchmark data explains why this check matters: on short tasks, the two models are within a fifth of a point. You may not notice which one responded until the work becomes long enough for Full’s advantage to appear.

Test the difference on your own work

If you have Muse Code Power, measure the gap with a task that reflects your real workload:

1. Choose one long task, such as a 30-step refactor or a multi-file code review.
2. Run it with --model muse-wonder-1.0.
3. Switch to Full through the /models menu.
4. Run the same task again.
5. Compare the finished result and the number of manual interventions.

That test tells you whether Full’s four-point overall advantage matters for your kind of work.

For the API side, use Apidog to verify that Wonder—not a benchmark headline—meets your quality bar. Store the API key as a secret variable, send your real prompts at medium and max reasoning effort, assert on the response shape, and inspect the usage field to calculate cost. The Muse Wonder API guide includes the requests. You can also [download and no third-party gateway. Meta has not announced a release date for API access.

Does maximum reasoning effort turn Wonder into Full?

No. Reasoning effort controls how long Wonder reasons. The model ID remains unchanged.

Is Full worth $50 per month if Wonder starts at $5?

For solo work on long coding tasks, the gaps are meaningful: 68 versus 63 on the Coding Agents Index and is the only version you can build with. Use%;">