Meta released Muse Wonder and Muse Wonder Full on October 12. Within a day, one group was pointing to frontier-level agent benchmarks while another was sharing underwhelming 3D-game screenshots. Neither is enough to decide whether to switch models. This guide separates access, memory, sustained work, independent evaluation, real tasks, cost, and permissions, then gives you a trial plan you can run on your own workload.
Which Muse Wonder model can you actually use?
Meta’s launch announcement draws a clear line. Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Wonder will later enter Muse experiences; “coming soon” is not current general availability. A demo shared by somebody else does not establish what your account can access.
| Item | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Launch access | Meta Model API; Muse Code | Muse Code subscription |
| API model ID | muse-wonder-1.0 on the pricing page |
No publicly listed per-token API SKU |
| Context window | 1M tokens in the independent evaluation | 1M tokens in the same evaluation |
| Modalities | Text, image, video, audio input; text output | Same, as recorded by the evaluator |
| Billing | API input, cached input, and output rates | Subscription access; do not apply Wonder’s API rates |
For programmatic deployments, first check the Meta Model API pricing and limits page and your own console. Full’s benchmark result does not imply that it is callable from your existing API key.
What Active Memory and Active Focus are meant to do
The Meta research post describes Active Memory as retaining useful context across conversations and retrieving it when relevant. Active Focus is meant to keep an agent moving through blockers, alternate approaches, and errors while seeking clarification and approval before important actions. These are product aims, not proof that every memory will be right or that sensitive actions should run unattended.
Consider a migration plan for a continuously deployed SaaS application. Test memory in three separate conversations: state the stack and a no-downtime constraint; later request a database migration; then suggest a step that conflicts with the earlier constraint. Did the assistant retrieve the right requirement, identify the conflict, and explain its source? For focus, give it a repository change that requires reading code, editing, running tests, fixing a failure, and reporting the diff. Repeatedly retrying a failing command without progress is not sustained execution.
Figure 1: Meta’s memory benchmark points to a test area; your team still needs to measure recall, false memories, and data boundaries.
Figure 2: SWE-Marathon targets long-running software work, but outcomes depend on the task set, tools, and stopping rules.
Permissions matter here. Check what conversations, calendars, and files the product can read; how memories can be removed; and when a tool action needs a human confirmation. Meta says it trained the models to avoid sharing private information without permission. That statement alone is not evidence that every deployment meets your organization’s data policy.
Muse Wonder benchmark gains are real within their test conditions
Artificial Analysis reports Intelligence Index v4.3 scores of 52 for Wonder and 56 for Full at maximum reasoning effort, versus 48 for Muse Spark 1.3. Full ranked in the top three on six of ten component evaluations. Its Coding Agents Index score was 68 in Muse Code, while Wonder scored 63; the same evaluator lists Claude Code with Sonnet 5.5 at 68 and with Opus 5.5 at 66.
Figure 3: The 52/56/48 figures share one evaluator, index version, and effort setting; they are not universal project success rates.
Figure 4: Meta’s own chart shows the claim it is making. Read it alongside independent tests without mixing their scoring rules.
Figure 5: Controlled legal and finance evaluations do not remove the need for source documents and qualified review in professional work.
The phrase “Full matches or exceeds Opus 5.5 in coding” needs qualifiers: these are model-plus-tool combinations with potentially different scaffolds, effort settings, and tasks. An index score of 68 is not “68% of projects delivered autonomously.” For production, error recovery, retries, human takeovers, and accepted outputs matter more than a single leaderboard position.
Figure 6: A social screenshot preserves only part of an evaluation; follow the full methods before making a procurement decision.
Why the strong scores and weak-looking 3D demos can coexist
Meta showed a playable kitchen game. The source GIF includes both an active scene and a completion screen, so this is more than a still promotional render. But without the full prompt, edit history, human intervention, and comparable control runs, it cannot establish a zero-shot success rate.
Figure 7: A completion screen shows one finished state, not that one prompt produced it without human edits.
Figure 8: The later scene and completion state differ; frame rate and input responsiveness still require hands-on testing.
Another creator placed a Muse Wonder Full mower game next to an Opus 5.5 result. At two points in the video, scene density, materials, and architectural detail visibly differ. This is one creator’s demonstration with no complete public record of prompts, tools, budget, iterations, or manual fixes. It raises a good test question—similar benchmark scores need not imply similar visual output—but cannot rank the models across all 3D projects.
Figure 9: The video labels are creator-reported project costs, not general model prices.
Figure 10: A second time point helps inspect interaction and scene changes; generation conditions may still differ.
The animation and rocket examples belong in the same category: useful prompts for your own tests, not controlled replication. If your use case is generated 3D game code, capture the engine, dependencies, exact prompt, number of runs, permitted patches, frame rate, and playable duration. A blind hands-on test by people who did not build the demos tells you more than one attractive screenshot.
Figure 11: One animation frame shows an output but cannot establish temporal consistency or the amount of human work.
Figure 12: The rocket comparison is anecdotal; read any cost or quality claim together with the full creation process.
Calculate accepted-task cost, not just the input-token rate
As of October 13, 2026, Meta’s official API rates for Muse Wonder standard mode are $4 per million uncached input tokens, $0.20 per million cached input tokens, and $20 per million output tokens. The model ID is muse-wonder-1.0. Contributor mode is unavailable for Wonder on the API. Built-in web search adds $2.50 per 1,000 queries. Full has no published per-token API price, so do not compare a subscription allowance to a single API call.
Here is arithmetic, not a measured invoice: 200,000 uncached input tokens, 800,000 cached input tokens, and 50,000 output tokens would cost 0.2×$4 + 0.8×$0.20 + 0.05×$20 = $1.96 in token charges. Search, retries, and human fixes are extra. Look at the actual cached_tokens usage field; do not assume all prior context qualifies for the cache rate.
Figure 13: Artificial Analysis measured about $4.52 per Wonder max-effort Intelligence Index task; that workload is not a standard chat bill.
The evaluator recorded roughly 7.69M input tokens and 54,000 output tokens per index task, with Wonder at $4.52 per task versus Spark 1.3 at $1.60. A higher-capability model may still be cheaper overall if it eliminates rework; a lower call price may be costly if many attempts fail. Use this accounting rule:
Total cost per accepted task = (attributable API/subscription cost + tool fees + review and rework) / number of accepted tasks
A one-week, 20-task migration trial
Start with 20 real, reproducible tasks: five short knowledge questions, five cross-file coding changes, five tasks that need information from a prior conversation, and five long-running tasks with an injected failure. Keep earlier inputs and outputs from your current model for each type. Sanitize sensitive data, and hold tool permissions, timeouts, and intervention rules constant.
| Measure | Record it this way | Red flag |
|---|---|---|
| Acceptance | Pass or fail against a written test | Trusting the model’s “done” message |
| Recovery | Inject a broken link, failing test, or interruption | Loops, skipped errors, fabricated success |
| Memory | Ask for an earlier constraint and introduce a conflict | Missed constraint or false private memory |
| Full cost | Tokens, search, retries, reviewer minutes | Comparing input rate alone |
| Control | Inspect file, network, and outbound actions | Broad permissions by default |
Use the first five tasks to calibrate the scoring rubric, then run the other fifteen. Retain model name and version, effort setting, prompt, tools, logs, usage fields, and final artifacts. For disputed results, ask two reviewers to score outputs without knowing which model produced them. Expand deployment only when accepted-task rate, elapsed time, or human rework improves consistently and cost and permissions stay within your limits.
Figure 14: Familiar and unfamiliar task results are a reminder to include your own long-tail work, not only popular demos.
A migration decision for ClaudeAPI readers
For teams evaluating models in the wider developer market served by ClaudeAPI, similar token prices are not evidence of equivalent outcomes. Establish a baseline for accepted tasks, output tokens, cache hits, retries, and reviewer minutes; then test sanitized copies through Meta’s official Wonder access. Full’s Muse Code subscription is a different access and purchasing unit from an API workflow, so track it separately for now.
Record the price snapshot date, currency, official pricing link, and formula. When rates or quotas change, you can rerun the calculation instead of letting an old screenshot drive the budget. This article does not imply ClaudeAPI has integrated Muse Wonder.
Sources and measurement notes
- Meta launch announcement: launch date, access, memory/focus, and safety claims.
- Meta research blog: technical framing, vendor evaluations, and demos.
- Meta Model API pricing and limits: Wonder’s model ID, API prices, cache, and extra charges.
- Artificial Analysis evaluation: independent index, coding-agent, and cost-per-task figures; the numbers may change as evaluation versions change.
- Original report and demo images: creator cases and screenshots. Screenshot costs and individual experiences are treated as anecdotes here.
FAQ
Can I call Muse Wonder Full through an API today?
As of October 13, 2026, Meta only confirms Full for Muse Code subscribers. Its public API pricing page lists muse-wonder-1.0. Check your own console and the current official docs before planning an integration.
Does a score of 56 prove that Full is better than Opus 5.5 at everything?
No. The 56 is Artificial Analysis Intelligence Index v4.3 at maximum reasoning effort. Results can change with your tasks, tools, budget, and amount of human repair.
Will Active Memory automatically read all my private data?
You cannot infer access scope from the feature name. Check permissions, connected data sources, retention and deletion controls. Use sanitized examples at work and require human approval for outbound or modifying actions.
What will one API task actually cost?
Multiply actual uncached input, cached input, and output token counts by the published rates, then add search, retries, and review. The $1.96 example is arithmetic for stated usage; $4.52 is an evaluator’s benchmark average. Neither is a fixed bill for your tasks.
Should an existing coding-agent team migrate immediately?
Run the same 20 tasks with the same tool permissions and acceptance tests first. Full’s result in Muse Code does not automatically transfer to your IDE, repository, or API setup.
