Muse Wonder
Model ranking
- Arena Score
- 1503.3155733721
- Arena reference rank
- 7
- Evaluation configuration
- Muse Wonder (Max)
- Source published
- 2026-11-19
- Votes
- 16158
- Confidence lower
- 1497.5609021519
- Confidence upper
- 1509.0702445924
- Last successful check
- 2026-11-22T14:00:05.877Z
- Last value/rank change
- 2026-11-22T09:30:20.053Z
- Methodology
- Human preference — Text / overall; checked every six hours.
Source: Arena / LMArena ↗ · CC BY 4.0. Selection and catalog matching by XPU LIVE.
| Provider | Price | Provider model ID | Status | Context / max output | Source |
|---|---|---|---|---|---|
Router catalog Tariff not listedInput $4USD / 1M tokensOutput $20USD / 1M tokens Updated 2026-11-22 · 15:43 UTCOpenRouter · Router catalogInput: $4 Output: $20 Cache Read: $0.2 Web Search: $0.01 OpenRouter route catalog price, not a direct-provider price. Routing, modality and endpoint conditions can change the actual bill. Checked 2026-11-22T15:43:15.988Z Price source ↗ | meta/muse-wonder | Catalog-listedConditionsOpenRouter catalog quote; upstream conditions and regional availability can differ. The pay-as-you-go Model API uses Standard mode for Wonder. Contributor mode is not available for that API model. Full is a separate subscription-only variant; neither variant is currently offered in South Korea according to Meta’s model documentation. | Context tokens 1048576Output tokens 943718 | Official source ↗2026-10-31 |
Benchmark review · November 19, 2026. The latest verified Artificial Analysis Intelligence Index score is 44 for Muse Wonder Max, checked on November 19, 2026. Rank is among verified models in this curated catalog. Benchmark source.
Muse Wonder is Meta’s frontier reasoning model for long-running coding and professional agent tasks. The regular model is available through Meta Model API and Muse Code, with a separate Full variant for selected subscriptions.
Muse Wonder specifications and access
The pay-as-you-go Model API uses Standard mode for Wonder. Contributor mode is not available for that API model. Full is a separate subscription-only variant; neither variant is currently offered in South Korea according to Meta’s model documentation.
The specification table distinguishes manufacturer references from provider deployment limits. Check the linked endpoint before choosing a maximum prompt or response size. If a field is unpublished, treat it as unknown rather than assuming the limit of another model in the family.
Muse Wonder pricing and cost planning
Meta’s Standard rate is $4 input, $20 output and $0.20 cached input per million tokens. Web search grounding adds $2.50 per 1,000 queries. The reviewed price page states there is no long-context premium.
For a practical budget, count input, generated output and repeated attempts separately. Also record tool charges and caching conditions. Compare several complete tasks with the same acceptance criteria. A low input price alone can hide expensive long answers or additional retries.
Where to use Muse Wonder
Compare multi-file software work, research synthesis and tool-driven workflows. Use representative tasks that include recovery from mistakes, because a successful first response does not measure a complete agent run.
Start with a small evaluation set containing easy, difficult and incomplete requests. Keep the expected result for each task. Then inspect accuracy, response time and the effort needed to correct failures. Use those observations to decide whether a more capable or less expensive route fits the workload.
Muse Wonder benchmark and alternatives
The current comparison uses the Artificial Analysis Intelligence Index v5.0 snapshot checked on October 31, 2026. The recorded configuration is max. Its score describes that evaluation setting; a different reasoning budget or tool setup can change the result.
Review the official Muse Wonder reference and the technical documentation. Use the benchmark evaluation to check the tested configuration.
Compare Muse Wonder with other entries in the AI model rankings and pricing directory. For an integration, inspect the linked provider offers and run your own acceptance tests before choosing a production endpoint.
Specifications
- Catalog model ID
- meta/muse-wonder
- Developer model ID / revision
- Muse Wonder
- Input modalities
- text, image, video, file, audio
- Output modalities
- text
- Context / input token limit
- 1,000,000
- Maximum output tokens
- 943,718
- Tool calling
- Supported
- Structured output
- Supported
- API parameters (source-specific)
- include_reasoning, max_tokens, reasoning, reasoning_effort, repetition_penalty, response_format, structured_outputs, temperature, tool_choice, tools, top_k, top_p
- Weight access
- Proprietary / hosted API
- Weight license / access terms
- Meta service terms; proprietary weights
- Total parameters (billions)
- Not publicly disclosed
- Active parameters (billions)
- Not publicly disclosed
- Download size (GB)
- Not applicable — no public weights
- Release date
- 2026-10-12
- Notes
- Meta model ID: muse-wonder-1.0. Meta publishes a 1M context; OpenRouter publishes 1,048,576 context and its own output limit. Weight size and parameter count are undisclosed. OpenRouter context, maximum output and supported parameters describe that route; they are not universal limits across every provider.
Catalog entries link to their specification and availability sources. This is a curated selection, not every model each provider offers. Best prices are the lowest collected USD tariffs, not a total-cost estimate. Token offers are selected by input price, with output from the same tariff; workload mix may change the cheapest provider. Listed variants may require batch processing, off-peak hours, or specific media settings. Price coverage varies by provider; collected tariffs include their source and observation time. Arena user-preference scores are checked automatically every six hours against the latest published Text / overall dataset. Source: Arena / LMArena; CC BY 4.0. Selection and catalog matching by XPU LIVE. They apply to the shown evaluation configuration; listed prices may use different settings. Models without verified scores are not assumed to be weaker.