Overview
Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026. This series gives developers a new choice for demanding coding and agent workloads beyond Muse Spark 1.3. Wonder is available through Meta Model API and Muse Code; Full is limited to eligible Muse Code subscriptions. The API overview still labels Wonder as preview. Launch announcement, API overview.
For an application you control, evaluate Wonder. For difficult work inside Muse Code, compare Wonder with Full before paying for a higher plan. This guide uses official documentation and reported evaluations; it does not claim hands-on testing.
What's New
A stronger model family beyond Spark
Wonder is a new model line, rather than a Spark 1.4 update. Meta describes larger base models and shows Wonder Full performing better on unfamiliar tasks than the Spark 1.3 trend predicts at a given level of performance on familiar tasks. This is a reason to evaluate broader workloads, not a promise of a fixed improvement on every project. Research introduction.
Active Memory and Active Focus
Active Memory carries relevant context across conversations and retrieves it for the current task. Active Focus describes persistence through blockers, alternative approaches and error recovery. Meta also describes clarification and checks before important actions. These are model behaviors to evaluate in an agent harness; the announcement does not establish a configurable memory API or a guarantee that every action needs approval. Research introduction.
A higher-capability Full variant with a different access route
Wonder Full introduces a separate model choice for Muse Code users on Power Usage or Maximum Usage. It is not an API reasoning setting. The release also introduces Maximum Usage, with 80x Everyday usage limits and five-hour and weekly resets. Developer release post.
Wonder vs Full
| Decision | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Own application or agent | Model API: muse-wonder-1.0 |
Not available on Model API |
| Muse Code access | Pay-as-you-go or any subscription plan | Power Usage or Maximum Usage subscription |
| Context | 1,048,576 tokens | Model card specifies 1M tokens |
| Reasoning versus identity | Setting effort to max still selects Wonder | Separate higher-capability variant |
| Best evaluation starting point | Your API integration or existing Muse Code tasks | Difficult Muse Code tasks that justify a higher plan |
Access and reasoning distinctions come from the model documentation; Full's context is listed in the model card.
Meta's scorecard reports DeepSWE 1.1 results of 78.8 for Wonder and 79.0 for Full, and SWE-Marathon results of 53.1 and 57.5 respectively. Full's SWE-Marathon result uses xhigh reasoning, despite the table's general max headings. The model card identifies third-party result sources and distinguishes them from controlled internal comparisons. Full is also slightly below Wonder on MemoryBench, so the published results do not support treating it as better at everything. Use them to shortlist, then compare completed tasks, corrections, latency and cost on your own workloads. Official scorecard, evaluation methodology.
Compared With Muse Spark 1.3
Spark 1.3 already supported long-running workflows, clarification and confirmation before consequential actions. Those behaviors should not be described as first introduced by Wonder. Spark 1.3 release post.
| Choice | Spark 1.3 | Wonder series |
|---|---|---|
| Workload position | Cost-sensitive or high-volume work | More complex, long-running agent tasks |
| API context | 1,048,576 tokens | Wonder: 1,048,576 tokens; no context-size increase |
| API pricing per 1M tokens | $1.25 input / $0.15 cached input / $4.25 output | Wonder: $4 input / $0.20 cached input / $20 output |
| Discounted API Contributor mode | Available for Spark 1.3 | Not available for Wonder |
| Higher-capability subscription choice | Existing Spark models | Full on eligible Muse Code plans |
Workload positioning comes from the model documentation; prices and API Contributor availability come from pricing documentation. Spark remains a reasonable default where its quality is sufficient and the higher Wonder token rates do not pay for themselves through better task completion.
Availability, Cost & Data Settings
Current developer access
Wonder's API Standard mode charges the rates above. The published per-team limits are 3,000 requests and 4,000,000 tokens per minute. Requests from all keys in a team share those limits. Web search grounding adds $2.50 per 1,000 search queries to token costs. API pricing and limits.
Full requires Power Usage, listed at $50/month, or Maximum Usage, listed at $200/month. Subscription access has usage limits; it does not provide unrestricted Full API access. Verify local eligibility and checkout terms before subscribing. Model and plan pricing.
Muse Code subscriptions default to Contributor mode, which permits Meta to use submitted content and responses to train future models in exchange for more usage. You can disable it in settings. Subscription benefits apply only to Muse Code's CLI credential; additional API keys are billed pay-as-you-go. Availability and benefits may vary by region. Subscription documentation.
Consumer Muse rollout remains a separate stage
The October 12 announcement says Wonder will power Muse later, starting with subscribers. It does not give a confirmed consumer rollout date. Developer availability should not be read as access for every Muse app user. Official announcement.
Migration & Evaluation
For an existing Spark API integration, keep the same request format and endpoint and change the model to muse-wonder-1.0. Meta documents https://api.meta.ai/v1 as the base URL. This request compatibility does not establish equivalent behavior, latency or total task cost. API overview.
Run representative repository changes and agent tasks before switching default traffic. Keep tool permissions and approval rules fixed, and compare task completion, failed tool calls, total tokens, wall-clock time and human intervention. Check your team's enabled model catalog and quota. Evaluate Full inside Muse Code separately; increasing Wonder's API reasoning effort will not grant Full.
Who Should Upgrade / Who Should Wait
- Evaluate Wonder if Spark 1.3 struggles with complex multi-file changes or agents that must recover from blockers, and you can accept preview access and higher token rates.
- Evaluate Full if you already work in Muse Code and task-level gains justify Power or Maximum subscription costs and limits.
- Keep Spark 1.3 if its results meet your quality bar and your workload depends on low cost, low latency or discounted API Contributor access.
- Wait if you require Full in your own API application, guaranteed consumer Muse availability, or production commitments beyond the currently documented preview.
Sources
Official evidence was checked on October 14, 2026. Prices, eligibility and access can change; confirm the linked documentation and your account before migrating.
