As of: 2026-10-12 22:30 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/m/muse-wonder/

# Muse Wonder

Meta, Muse family. Status: preview. Type: reasoning model. Released: 2026-10-12. Training cutoff: not published.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Meta Model API | `muse-wonder-1.0` | https://api.meta.ai/v1/chat/completions | https://dev.meta.ai/docs/models |
| OpenRouter | `meta/muse-wonder` |  | https://openrouter.ai/meta/muse-wonder |
| Muse Code (every plan and pay-as-you-go) |  |  | https://dev.meta.ai/products/muse-code |

## Pricing

- Input: $4 per 1M tokens
- Cached input: $0.20 per 1M tokens
- Output: $20 per 1M tokens
- Unit as published: per 1M tokens (USD), Standard mode; no Contributor mode on pay-as-you-go
- Source: https://x.com/artificialanlys/status/2109697083519447118
- Last checked: 2026-10-12

## Specs

- Input: text, image, video, audio, pdf
- Output: text
- Context window: 1,048,576 tokens
- Max output: not published
- Open weights: no
- Licence: proprietary

## What it has not seen

- No published training cutoff. Counted from its release (2026-10-12): at least 0 AI events since (0 major, 0 historic), as of 2026-10-12.
- Everything since September 2026: https://postcutoff.com/since/2026-09/

## What stands out

- Active Memory: Carries important context about the user across conversations and retrieves it when relevant; 85.6 on Meta's internal MemoryBench versus 81.5 for Claude Opus 5.5 (Meta's own run). Source: https://research.meta.ai/blog/muse-wonder
- Active Focus: Keeps working toward a goal through blockers, tries alternative approaches, asks for clarification and checks in before important actions. Source: https://research.meta.ai/blog/muse-wonder
- Long-horizon coding: DeepSWE 1.1 78.8 and SWE-Marathon 53.1 in Meta's table, above Claude Opus 5.5 (74.2, 46.3) and GPT-6 Astra (74.1, 42.5); plans, delegates to sub-agents and calls tools in parallel. Source: https://research.meta.ai/static/muse-wonder-model-card
- Computer use: Operates browsers, desktop and mobile apps; OSWorld V2.1 72.6 partial / 34.3 binary (Meta's run). Source: https://dev.meta.ai/models/muse-wonder/

## Notes

Docs say 'in preview'. Reasoning effort 'minimal' to 'max'. Same endpoints as Muse Spark (Responses, Chat Completions, Messages), so switching is a model-id change. Artificial Analysis Intelligence Index v4.3: 52 at max; $4.52 per index task. Prices come from Artificial Analysis, TechCrunch and the OpenRouter listing; Meta's own pricing page was not read. Knowledge cutoff and max output not published. The larger Muse Wonder Full is a separate model (muse-wonder-full).

Meta's first frontier-class API model (internal codename "Watermelon" per CNBC), released Oct 12, 2026.

```bash
curl https://api.meta.ai/v1/chat/completions -H "Authorization: Bearer $MODEL_API_KEY" -H "Content-Type: application/json" \
 -d '{"model":"muse-wonder-1.0","messages":[{"role":"user","content":"Hello"}]}'
```

Sources: https://dev.meta.ai/docs/models , https://dev.meta.ai/models/muse-wonder/ , https://research.meta.ai/static/muse-wonder-model-card , https://openrouter.ai/api/v1/models

Model page: https://postcutoff.com/m/muse-wonder/