What Is Meta Muse Wonder? Official Pricing, 1M Context, Benchmarks, and Hands-On Comparison with Claude Opus 5.5

Meta released Muse Wonder on October 12, alongside Muse Wonder Full, which is only available through a Muse Code subscription. The API version, muse-wonder-1.0, is officially priced at $4 per million input tokens and $20 per million output tokens, the same as Claude Opus 5.5, with a 1M context window. This post covers how the two versions differ, official and third-party benchmarks, the strengths and weaknesses seen in hands-on testing, and how to call it.

8 分钟阅读
What Is Meta Muse Wonder? Official Pricing, 1M Context, Benchmarks, and Hands-On Comparison with Claude Opus 5.5

Bottom line: Meta Muse Wonder is a closed-source flagship model Meta released on October 12, 2026, focused on coding, long-running agents, and computer use. The API model name is muse-wonder-1.0, officially priced at $4 per million input tokens, $20 per million output tokens, and $0.20 for cached input, with a context window of 1,048,576 tokens.

Two things to get straight first: the best-looking scores at Meta's launch event belong to Muse Wonder Full, which is only available in the Muse Code subscription and can't be called through the API. The Muse Wonder you can use through the API scores 52 on the Artificial Analysis Intelligence Index. It also thinks a lot, and several users found in testing that it burns more tokens and runs slower than Claude Opus 5.5 at the same price.

Figures in this article are current as of 2026-10-21, based on Meta's official research blog, developer blog, model page, pricing docs, plus Artificial Analysis, OpenRouter and other third-party pages. Third-party leaderboard scores change, so defer to the original pages.

What Is Muse Wonder

What Meta's official blog says:

  • Released on October 12, 2026, with two versions, Muse Wonder and Muse Wonder Full, positioned as agent models for personal and work use
  • Trained in Meta's Prometheus data center, built on a larger base model than Muse Spark
  • Two headline capabilities:Active Memory, which remembers important context across conversations and recalls it when needed;Active Focus, which switches approaches when a task stalls, asks clarifying questions when needed, and confirms before important actions
  • Muse Wonder is live on the Meta Model API and Muse Code, and can also be called on OpenRouter (model name meta/muse-wonder)
  • The developer docs currently say that Muse Wonder is still in preview (preview)

How the two versions differ:

VersionWhere to use itPricing
Muse Wonder (muse-wonder-1.0)Meta Model API, Muse Code, OpenRouterBilled per token, $4 / $20
Muse Wonder Full (muse-wonder-full-1.0)Muse Code onlySubscription only: Power Usage $50/month, Maximum Usage $200/month

The official docs explicitly note: On the API, setting the reasoning level of muse-wonder-1.0 to max does not turn it into Muse Wonder Full.

Official Pricing

Meta Model API pricing docs, per million tokens, in US dollars:

muse-wonder-1.0Price
Input$4.00
Cached input$0.20
Output$20.00

Billing details:

  • No long-context surcharge, the same unit price whether you use a little context or a lot
  • Reasoning tokens are billed at the output rate and listed separately in the returned usage
  • Using the built-in web_search tool costs an additional $2.50 per 1,000 queries
  • Meta's Contributor mode (lets you trade training on your data for lower prices)does not apply to Muse Wonder, only to Muse Spark

This price matches Claude Opus 5.5 at $4 / $20. But the same price doesn't mean the same cost, as the hands-on section below explains.

Once AI Code With lists it, we'll add our pricing here.

Context and Capabilities

ItemOfficial description
Model namemuse-wonder-1.0
Base URLhttps://api.meta.ai/v1
Context1,048,576 tokens
InputText, images, video, PDF, and more; confirm your channel supports audio first
OutputText
Reasoning levelminimal / low / medium / high / xhigh / max, chosen by the model itself when omitted
ProtocolsResponses API, Chat Completions, Anthropic Messages compatible endpoints
Parameter countNot disclosed
The input modalities aren't described consistently: OpenRouter lists text, images, video, audio, and PDF; Artificial Analysis lists text, images, speech, and video; and Meta's developer blog cautions that "audio-dependent scenarios require confirming whether the release channel supports it." If you need audio, test with a small sample first.

A few reasoning gotchas, all from the official docs:

  • Thinking can't be turned off.reasoning_effort set to none returns HTTP 400
  • The output cap includes thinking tokens.max_tokens (Chat Completions) or max_output_tokens (Responses API) set too low, the model may hit the cap after thinking and truncate the answer
  • For multi-turn agents, use the Responses API. With external API keys, Chat Completions doesn't preserve reasoning content from the previous turn, so every turn starts thinking from scratch; the Responses API can continue from the previous turn using previous_response_id or encrypted reasoning content

Reviews and Benchmarks

Official Data

All of the following were measured by Meta itself, with comparison models run at their highest reasoning level (Muse Wonder Full's SWE-Marathon run uses the xhigh level) and no third-party verification:

BenchmarkWhat it measuresMuse WonderMuse Wonder FullClaude Opus 5.5GPT-6 Astra
DeepSWE 1.1Long-task coding78.879.074.274.1
SWE-MarathonUltra-long-task engineering53.157.546.342.5
Terminal-Bench 4.0-AAAgentic coding in the terminal52.057.159.659.1
ProgramBenchWriting programs from scratch50.059.070.552.5
MemoryBenchRemembering context85.685.581.581.4
AutomationBench-AAEnd-to-end business workflows39.145.441.241.6
Artificial Analysis Index (from Meta's table)Overall52565853

How to read this table:

  • Long-task coding (DeepSWE, SWE-Marathon) and memory (MemoryBench) are Muse Wonder's strengths, and the API version also scores above Opus 5.5 on these
  • On Terminal-Bench 4.0 and ProgramBench, the API version of Muse Wonder trails Opus 5.5 by a clear margin
  • On the overall index, Opus 5.5 still ranks first in Meta's own table (58), Muse Wonder Full scores 56, and the API version scores 52

Third-Party Data

Artificial Analysis(as of 2026-10-21, original page):

  • Intelligence Index 52, ranked 9th out of 214 models at a similar price
  • Output speed 152.2 tokens/s (Meta's official API)
  • First-token latency 31.59 seconds (including thinking time), which suggests it often thinks for a long while before it starts answering
  • Running one index task takes an average of $4.52; running the full index produces 130M tokens, versus a median of 82M in the same tier, so it leans wordy

Vals AI (reposted on the OpenRouter model page, as of 2026-10-21):

  • Vals Index 64.9%
  • Terminal-Bench 4.0 49.5%, SWE-bench Multimodal 45.6%, Vibe Code Bench v1.1 83.1%
  • Finance Agent v2.1 65.5%

Design Arena (also reposted on that page): website generation 1361, game development 1379, UI components 1371 Elo.

OpenRouter measurements: Over the past week, P50 throughput was 84 tokens/s and P50 latency was 3.03 seconds. These differ from AA's numbers because the test methods and request types are different.

Community testing

  • Same price, but costlier: X user @eddieor compared Muse Wonder, Claude Opus 5.5, and GPT-6.1 Sol on three everyday tasks (a single-file Pomodoro timer, a function with 6 planted bugs, and a roast based on real source material). They ran through Vercel AI Gateway with default reasoning settings and scored the results with 14 Playwright tests and 6 unit tests. All three passed. But total time and cost differed a lot, original post:
ModelTotal timeOutput tokensCost
Muse Wonder73.8 seconds8,670$0.181
Claude Opus 5.539.4 seconds4,494$0.101
GPT-6.1 Sol93.2 seconds4,339$0.048

The unit price is the same, yet Muse Wonder cost about 80% more than Opus 5.5 because it output nearly twice as many tokens. On the Pomodoro timer alone it output 5,968 tokens, 3,577 of which were thinking.

  • Overthinks even on the low tier:VulcanBench analyzed the run logs and found that Muse Wonder overthinks even on the low tier, taking more than 4 times as long as GPT-6.1 Sol to finish the same tasks
  • Strong at agents, weak at frontend: On the How I AI podcast from Lenny's Newsletter, host Claire tested it on nearly 20 real tasks. Muse Wonder ranked first among models at its price that she has tested on billing tickets, inbox triage, computer use, and email writing, and in the billing task it even caught a prompt injection. Frontend design was a weak spot: in the redesign task it ignored brand guidelines, scattered emoji around, and broke images. Prototypes got about 80% of the skeleton right, but in the interaction stage buttons often failed and themes came out wrong (Chinese write-up)
  • Solid on finance and accounting tasks: In DualEntry's accounting AI benchmark, Muse Wonder ranked 2nd overall, 21.8 points above Muse Spark 1.1 (original post)

All of the above are tests by individuals or organizations, with different tasks and environments, so treat them as a rough reference only. We'll update this once more third-party evaluations come out.

Pros, cons, and best-fit scenarios

Pros:

  • Long-running coding tasks and multi-step agents are its strengths: it leads on the official long-task benchmarks, and third-party agent testing has been positive too
  • The 1M context window comes at no extra charge, and cached input costs just $0.20
  • Highest score on memory tasks (MemoryBench)
  • Compatible with both OpenAI and Anthropic request formats, so migration cost is low

Cons:

  • It thinks a lot and is slow to the first token, so at the same price its real-world spend is often higher than Opus 5.5
  • The strongest model at the launch event was Muse Wonder Full, which isn't available through the API
  • It trails Opus 5.5 on terminal and from-scratch programming tasks like Terminal-Bench and ProgramBench
  • Testers explicitly marked down its frontend and visual design
  • Still in preview

Good fit: long-running agent workflows, assistants that need to remember context across many turns, business automation like ticket handling and inbox triage, and long coding tasks in large repos.

Not a great fit: frontend pages and UI design, latency-sensitive interactive scenarios, and small tasks where per-task cost matters.

How to connect

Once it's listed on AI Code With, we'll add our model name and Claude Code and Codex configuration examples here.

If you want to use Meta's official API right now: get an API key at dev.meta.ai, then change the Base URL of your OpenAI-compatible client to https://api.meta.ai/v1 and set the model name to muse-wonder-1.0. The minimal request from the official docs looks like this (Responses API):

bash
curl -X POST "https://api.meta.ai/v1/responses" \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "muse-wonder-1.0", "reasoning": {"effort": "high"}, "input": "你的问题"}'

A few usage tips:

  • Use low or medium for everyday small tasks, and save max for the genuinely hard problems, otherwise cost and wait time climb noticeably
  • Set a generous output limit to leave enough room for thinking tokens
  • Use the Responses API for multi-turn agents

For general setup with third-party APIs, see "How to Connect Claude Code to a Third-Party API" and "Codex config.toml Configuration Guide".

FAQ

What's the difference between Muse Wonder and Muse Wonder Full?

Full is stronger, but it's only available in Muse Code's Power Usage ($50/month) and Maximum Usage ($200/month) subscriptions. The API only has muse-wonder-1.0, and even if you raise it to the max tier, it isn't Full.

Muse Wonder or Claude Opus 5.5?

The unit price is the same. In Meta's official table, Muse Wonder scores higher on long-task coding and memory, while Opus 5.5 scores higher on Terminal-Bench, ProgramBench, and the composite index. In hands-on tests Muse Wonder used more tokens, so the same task often cost more and took longer. Run your own tasks through both before deciding.

Is Muse Wonder open source?

No, and the parameter count hasn't been published. The model Meta open-sourced at the same time is Muse Spark Open.

Can I turn off thinking?

No.reasoning_effort set to none returns a 400. To save money, drop it to minimal or low.

Sources (checked 2026-10-21)

相关文章

Tencent Hunyuan Hy4 Released: Pricing, 1M Context, Benchmarks, and How to Use It in Claude Code / Codex

Tencent Hunyuan Hy4 Released: Pricing, 1M Context, Benchmarks, and How to Use It in Claude Code / Codex

Tencent Hunyuan released Hy4 on November 4. It adds vision and computer use for the first time, has a 1M context window and up to 64K output, and is priced at $0.834 input, $2.501 output and $0.042 cached input per million tokens, with a 50% launch discount on OpenRouter. This guide covers the official specs, pricing, the benchmark numbers you can check in Tencent's own table, how to call the Hy4 API on Tencent Cloud TokenHub and OpenRouter (model IDs hy4 / tencent/hy4), and tested steps for using Chinese models in Claude Code and Codex.

LongCat-2.5 是什么?美团 1.6T 多模态 Agent 模型:官方价格 $0.30/$1.20、1M 上下文、基准与实测测评

LongCat-2.5 是什么?美团 1.6T 多模态 Agent 模型:官方价格 $0.30/$1.20、1M 上下文、基准与实测测评

美团在 11 月 1 日发布 LongCat-2.5,11 月 2 日开放 API:1.6T 总参数、约 48B 激活的原生多模态 MoE,主打长程 Agent 编程和视觉编程,1M 上下文,限时价每百万 tokens 输入 $0.30、输出 $1.20。本文把官方基准和第三方实测分开整理,讲清它强在哪、哪里还不稳,以及怎么接入 Claude Code 等 Agent 工具。