Meta发布Muse Wonder与Full模型，主打Agent长程任务
Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It
Meta最新旗舰Agent模型发布，明确针对长程任务并给出详细API接入方式与定价，开发者可直接评估其能力边界与集成成本。
Meta has released Muse Wonder and Muse Wonder Full, its most capable models so far. Muse Wonder is a frontier model from Meta built for personal and professional AI agents. Both models target long-running agent work: coding, legal and finance tasks, and computer use. Muse Wonder is live now on the Meta Model API and Muse Code. Muse Wonder Full is limited to paid Muse Code plans.
Meta 发布了 Muse Wonder 和 Muse Wonder Full，这是迄今为止其最强大的模型。Muse Wonder 是 Meta 构建的前沿模型，专为个人和专业 AI 智能体设计。这两个模型都针对长期运行的智能体工作：编码、法律和财务任务以及计算机操作。Muse Wonder 现已在 Meta Model API 和 Muse Code 上上线。Muse Wonder Full 仅限于付费的 Muse Code 计划。
TL;DR
TL;DR（太长不看）
- Size: 1M-token context, input of text, image, video and audio, output of text.
- Runs on: Meta’s cloud only, via the API or Muse Code CLI (macOS, Linux, Windows).
- Performance: Muse Wonder Full leads 6 of 14 rows in Meta’s benchmark table. It trails Claude Opus 5.5 on several.
- Best: Muse Wonder scores 85.6 on MemoryBench, the top result among all 5 models listed.
- Worst: Muse Wonder scores 50.0 on ProgramBench, against 70.5 for Claude Opus 5.5.
- 规模：100 万 token 上下文，输入文本、图像、视频和音频，输出文本。
- 运行平台：仅限 Meta 云端，通过 API 或 Muse Code CLI（macOS、Linux、Windows）运行。
- 性能：在 Meta 的基准测试表中，Muse Wonder Full 在 14 项指标中的 6 项领先。它在几项指标上落后于 Claude Opus 5.5。
- 最佳表现：Muse Wonder 在 MemoryBench 上得分 85.6，在所有列出的 5 个模型中排名第一。
- 最差表现：Muse Wonder 在 ProgramBench 上得分 50.0，而 Claude Opus 5.5 得分为 70.5。
What is Muse Wonder?
什么是 Muse Wonder？
Muse Wonder follows Muse Spark on Meta’s scaling ladder. Meta team states that it was trained at its Prometheus data center, with gains across pretraining and reinforcement learning. Meta credits a larger base model for better results on unfamiliar tasks.
Muse Wonder 紧随 Muse Spark 之后，位于 Meta 的扩展阶梯上。Meta 团队表示，它是在其 Prometheus 数据中心训练的，在预训练和强化学习方面均有提升。Meta 归功于更大的基础模型带来了在不熟悉任务上的更好结果。
Two named behaviors define the release:
此次发布定义了两种命名行为：
- Active Memory: carries user context across conversations and pulls back the details a task needs.
- Active Focus: keeps working on hard tasks, tries alternatives, and checks in before important actions.
- 主动记忆（Active Memory）：跨对话携带用户上下文，并提取任务所需的细节。
- 主动聚焦（Active Focus）：持续处理困难任务，尝试替代方案，并在执行重要操作前进行检查。
For developers, Muse Wonder plans long tasks and delegates to sub-agents. It also operates browsers, desktop and mobile apps.
对于开发者而言，Muse Wonder 会规划长任务并委派给子智能体。它还操作浏览器、桌面和移动应用。
How does Muse Wonder perform on benchmarks?
Muse Wonder 在基准测试中的表现如何？
All scores below come from Meta’s published table at max reasoning effort. Several rows use third-party sources such as Artificial Analysis, Vals AI and Datacurve.
以下所有分数均来自 Meta 发布的表格，且处于最大推理努力水平。其中几项指标使用了第三方来源，如 Artificial Analysis、Vals AI 和 Datacurve。
| Benchmark | Muse Wonder Full | Muse Wonder | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|
| Personal Superintelligence Index (internal) | 59.1 | 54.3 | 49.9 | 54.5 | 55.2 |
| Harvey LAB-AA | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Vals Finance Agent V2 | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| AutomationBench-AA | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| OSWorld V2.1 (partial) | 78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
| MemoryBench (internal) | 85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
| DeepSWE 1.1 | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| SWE-Marathon | 57.5 (xhigh) | 53.1 | 45.6 | 42.5 | 46.3 |
| ProgramBench | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
| AA Intelligence Index v4.3 | 56 | 52 | 53 | 53 | 58 |
| 基准测试 | Muse Wonder Full | Muse Wonder | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|---|
| 个人超级智能指数（内部） | 59.1 | 54.3 | 49.9 | 54.5 | 55.2 |
| Harvey LAB-AA | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Vals Finance Agent V2 | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| AutomationBench-AA | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| OSWorld V2.1（部分） | 78.4 | 72.6 | 80.7 | 75.8 | 81.8 |
| MemoryBench（内部） | 85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
| DeepSWE 1.1 | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| SWE-Marathon | 57.5 (xhigh) | 53.1 | 45.6 | 42.5 | 46.3 |
| ProgramBench | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
| AA Intelligence Index v4.3 | 56 | 52 | 53 | 53 | 58 |
Note: PSI, MemoryBench and FocusBench are internal Meta evaluations. The model card says competitor runs are best-effort and may not be tuned. SWE-Marathon results for rivals used Codex and Claude Code, so it is not a common-harness comparison.
注意：PSI、MemoryBench 和 FocusBench 是 Meta 的内部评估。模型卡片指出，竞争对手的运行结果尽力而为，可能未经过调优。SWE-Marathon 中对手的结果使用了 Codex 和 Claude Code，因此并非基于相同基准的比较。
Muse Wonder vs Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra
Muse Wonder 对比 Claude Opus 5.5、Claude Fable 5.1 和 GPT-6 Astra
| Feature | Muse Wonder | Muse Wonder Full | Claude Opus 5.5 | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|---|
| Developer | Meta | Meta | Anthropic | Anthropic | OpenAI |
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Context | 1M | 1M | 1M | 1M | 1.05M |
| Input modalities | Text, image, video, audio | Text, image, video, audio | Text, image | Text, image | Text, image |
| Open weights | No | No | No | No | No |
| API price (in / out per 1M) | $4 / $20 | Subscription only | $4 / $20 | $10 / $50 | $10 / $50 |
| Access | Model API, Muse Code | Muse Code Power ($50) or Maximum ($200) | Claude API | Claude API | OpenAI API |
| DeepSWE 1.1 | 78.8 | 79.0 | 74.2 | 67.4 | 74.1 |
| 特性 | Muse Wonder | Muse Wonder Full | Claude Opus 5.5 | Claude Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|---|
| 开发者 | Meta | Meta | Anthropic | Anthropic | OpenAI |
| 参数量 | 未披露 | 未披露 | 未披露 | 未披露 | 未披露 |
| 上下文窗口 | 1M | 1M | 1M | 1M | 1.05M |
| 输入模态 | 文本、图像、视频、音频 | 文本、图像、视频、音频 | 文本、图像 | 文本、图像 | 文本、图像 |
| 开源权重 | 否 | 否 | 否 | 否 | 否 |
| API 价格（每百万 token 输入/输出） | $4 / $20 | 仅限订阅 | $4 / $20 | $10 / $50 | $10 / $50 |
| 访问方式 | Model API、Muse Code | Muse Code Power ($50) 或 Maximum ($200) | Claude API | Claude API | OpenAI API |
| DeepSWE 1.1 | 78.8 | 79.0 | 74.2 | 67.4 | 74.1 |
How to use Muse Wonder: step-by-step for new users
如何使用 Muse Wonder：新用户分步指南
Step 1: Open dev.meta.ai. Click Start building to sign in. The Muse Code install command sits on the same page.
步骤 1：打开 dev.meta.ai。点击“开始构建”以登录。Muse Code 安装命令位于同一页面。
Step 2: Pick your model. Muse Wonder costs $4 input, $0.20 cached and $20 output per million tokens. Muse Wonder Full needs a Muse Code plan: Power at $50 or Maximum at $200 per month.
步骤 2：选择你的模型。Muse Wonder 的价格为每百万 token 输入 $4、缓存输入 $0.20、输出 $20。Muse Wonder Full 需要 Muse Code 计划：Power 版每月 $50 或 Maximum 版每月 $200。
Step 3: Sign in with your Meta account. Confirm your name, accept the Model API terms, and click Get started.
步骤 3：使用你的 Meta 账户登录。确认你的名字，接受 Model API 条款，然后点击“开始”。
Step 4. Add billing, then create an API key. New accounts see a “No payment method” banner, and Create API key stays disabled. Add a card under Billing, then open API keys and select Create API key. Export it as MODEL_API_KEY, per the quickstart.
步骤 4：添加账单信息，然后创建 API 密钥。新账户会看到“无支付方式”横幅，且“创建 API 密钥”按钮处于禁用状态。在“账单”下添加银行卡，然后打开“API 密钥”并选择“创建 API 密钥”。按照快速入门指南将其导出为 MODEL_API_KEY。
Step 5. Test it in the Playground. Open Chat, pick muse-wonder-1.0, and set temperature, max tokens and reasoning effort.
步骤 5：在 Playground 中测试。打开 Chat，选择 muse-wonder-1.0，并设置温度、最大 token 数和推理力度。
Step 6. Copy working code. Click Get code for a curl or Python snippet. It calls the Responses endpoint at https://api.meta.ai/v1/responses.
步骤 6：复制可运行的代码。点击“获取代码”以获得 curl 或 Python 片段。它调用 https://api.meta.ai/v1/responses 上的 Responses 端点。
The API is OpenAI-compatible, so the OpenAI SDK works with a new base URL:
该 API 与 OpenAI 兼容，因此 OpenAI SDK 可与新的基础 URL 配合使用：
import os
from openai import OpenAI
client = OpenAI(base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"])
resp = client.responses.create(model="muse-wonder-1.0", input="List 3 risks in this lease clause: ...")
print(resp.output_text)Step 7: Or use Muse Code in your terminal. On Mac, install with curl -fsSL https://dev.meta.ai/install.sh | bash. On Windows, use irm https://dev.meta.ai/install.ps1 | iex. Run muse in your project folder, trust the workspace, and sign in. The default model is muse-spark-1.2. Switch with /models, or launch with muse --model muse-wonder-1.0.
步骤 7：或者在终端中使用 Muse Code。在 Mac 上，使用 curl -fsSL https://dev.meta.ai/install.sh | bash 进行安装。在 Windows 上，使用 irm https://dev.meta.ai/install.ps1 | iex。在项目文件夹中运行 muse，信任工作区并登录。默认模型是 muse-spark-1.2。使用 /models 切换，或使用 muse --model muse-wonder-1.0 启动。
Key Takeaways
关键要点
- Muse Wonder ships at $4/$20 per million tokens with 1M context.
- Muse Wonder Full leads DeepSWE (79.0) and SWE-Marathon (57.5).
- Claude Opus 5.5 still leads ProgramBench, OSWorld and Terminal-Bench.
- Wonder Full is Muse Code only, from $50 per month.
- No parameters, open weights or full safety report yet.
- Muse Wonder 的定价为每百万 token 输入 $4/输出 $20，支持 1M 上下文窗口。
- Muse Wonder Full 在 DeepSWE（79.0）和 SWE-Marathon（57.5）上领先。
- Claude Opus 5.5 仍在 ProgramBench、OSWorld 和 Terminal-Bench 上保持领先。
- Wonder Full 仅为 Muse Code，每月 50 美元起。
- 目前尚无参数量、开放权重或完整的安全报告。
Check out the research blog, model card and developer docs. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
请查看研究博客、模型卡片和开发者文档。所有功劳归于本项目的研究人员。此外，欢迎在 Twitter 上关注我们，别忘了加入我们拥有 15 万+成员的 ML SubReddit 并订阅我们的新闻通讯。等等！你在 Telegram 上吗？现在你也可以加入我们的 Telegram 群组了。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力