最后更新:11/11/2026最后核验:2026-11-11
Mistral Studio 是什么
Mistral Studio 是 Mistral AI 面向开发者推出的平台,用于在 Mistral API 之上构建、部署和治理智能体、工作流与应用。文档把它描述为开发者控制台加 API:创建密钥、在 Playground 中做原型、交付智能体与工作流,并监控用量。产品页强调企业级隐私、安全和数据完全归属,文档则说明智能体和工作流可运行在 Mistral 云或自有基础设施上,而应用运行在 Mistral Cloud。
核心功能
- 智能体与工作流:智能体依据指令、上下文和工具推理每一步;工作流是确定性代码,由 Studio 保存状态并重试失败步骤
- 数据连接器:内置连接器和自定义连接器可接入企业数据源,并在智能体和工作流之间复用
- 评估工具:experiments、campaigns、judges 和 datasets;文档说明 tracing 与评估仅对企业版组织开放
- 治理:运行时 guardrails、产品页称对每次响应生效的 moderation,以及工作区、角色和审计控制
- 部署:智能体和工作流支持混合部署、专属部署和自托管部署;应用运行在 Mistral Cloud
- 开发者控制台:可设置有效期的 API 密钥、用于测试提示词的 Playground,以及用量监控
- 开发工具:agents 和 workflows Python SDK(Python 3.12 至 3.14),以及 Mistral CLI
- AI 注册表:统一编目模型、智能体、提示词、skills、datasets 和工作流,并提供血缘信息
- 模型访问:文本、OCR 和语音模型通过同一套 API 调用
最适合
定价
mistral.ai/pricing 上没有单独的 Studio 订阅。该页(2026-11-11 核实),该页销售的是 Vibe 套餐——Free、$14.99/月的 Pro、$24.99/用户/月的 Team,以及需联系的 Enterprise;定价页 FAQ 说明 Free 套餐同时提供 Mistral Studio 访问权限,因此免费档是 Vibe 与 Studio 共享的额度,而不是公开的 Studio 配额。模型用量单独计量:API 价格表显示 Mistral Large 4 输入 $0.7/百万 token、输出 $2.1/百万 token,Mistral Small 4 为 $0.15/$0.6,OCR 4.1 为每 1,000 页 $4,Voxtral Mini Transcribe 3 为每分钟 $0.0035,Voxtral TTS 3 为每百万字符 $13。同一页面的 FAQ 仍写着 Mistral Large 输入 $0.5/百万、输出 $1.5/百万,与 API 价格表不一致,做预算前应先确认实际计费费率。
优缺点
优点
- 在一个控制台中完成原型、开发、评估和治理
- 工作流可保存状态并重试失败步骤
- 为企业数据提供内置连接器和自定义连接器
- Python SDK、CLI 和 Playground 覆盖不同开发方式
- 文本、OCR 和语音模型共用一套 API
- 智能体和工作流既可在 Mistral 云上运行,也可运行在自有基础设施上
缺点
- 没有独立的 Studio 价格:模型用量按 token、页数或分钟计量,公开的 Vibe 套餐才是入口
- 按文档说明,可观测性和评估仅对企业版开放
- 不同工作负载的部署选项不同:应用运行在 Mistral Cloud,智能体和工作流的部署选择更多
- 文档建议新项目使用 Agents SDK,并指出旧的 Agents API 指南已废弃、Workflows 中的 durable agents 即将废弃
- 文档描述的是代码优先的平台,不应默认它提供无代码应用构建能力
- 核实范围:基于 2026-11-11 核实的 Studio 产品页、定价页、Studio 与平台概览文档以及 API 价格表;未创建账号,也未部署智能体
替代工具
通过单一 API 统一路由多家厂商的模型,而不是使用单一厂商的智能体、工作流和模型体系。
面向业务工具的无代码智能体与工作流构建器,而不是代码优先的 API 控制台。
托管型智能体基础设施,可跨多种编程智能体运行框架部署,而不局限于单一厂商的 Studio 平台。
常见问题
详情
平台
功能特性
- Build agents and workflows that persist state, retry on failure and resume automatically, with built-in and custom connectors to company data sources
- Iterate with experiments, campaigns, judges and datasets; govern with observability, guardrails and moderation running on every response
- Deploy hybrid, dedicated or self-hosted, and run agents and workflows on Mistral’s cloud or your own infrastructure (apps run on Mistral Cloud)
- Console-side tooling: API keys with expiry, Playground, agents/workflows Python SDKs, the Mistral CLI, an AI registry with lineage, plus text, OCR and speech models behind the same API
支持语言
已知限制
- https://mistral.ai/pricing/ (2026-11-11) sells Vibe plans, not a separate Studio subscription: Free (limited messages, web searches and coding sessions, basic image generation, 100+ connectors), Pro at $14.99/month (with a $14.99 Education plan for verified students and a $5.99/month student price stated in the FAQ), Team at $24.99/user/month, and Enterprise as “Contact us”. Its FAQ says the Free plan also gives access to Mistral Studio, so the free tier is a shared Vibe/Studio allowance rather than a published Studio quota
- Model usage is billed separately and per unit. On https://docs.mistral.ai/inference/pricing (2026-11-11, Standard tab): Mistral Large 4 $0.7/M input, $0.07/M cached, $2.1/M output; Mistral Small 4 $0.15/$0.015/$0.6; OCR 4.1 $4 and OCR 3 $2 per 1,000 pages; Voxtral Mini Transcribe 3 $0.0035/min; Voxtral TTS 3 $13/M characters; Codestral $0.3/$0.03/$0.9; Voice agents $0.04/min on Mistral Small 4 and $0.08/min on larger models. Batch and Priority tabs exist on the same page. The pricing-page FAQ still quotes “Mistral Large costs $0.5 /M tokens in and $1.5 /M tokens out”, which does not match the $0.7/$2.1 shown for Mistral Large 4 in the API price table; both figures are recorded here and the conflict was not resolved
- The docs page states that agents should be built with the Agents SDK, that the older Agents API guides have moved to a Deprecated section, and that “durable agents in Workflows are about to be deprecated”. The Agents and Workflows SDKs are Python packages requiring Python >=3.12,<3.15, and the loop assumes you bring your own key stored in MISTRAL_API_KEY
- Per the docs, tracing and evaluating each change on real cases is an Enterprise-tier capability. The product page lists Hybrid, Dedicated and Self-hosted deployment, but the pricing page routes private deployments to Enterprise “Contact us”, so self-hosting is not self-serve. This listing is based on the Studio product page, the pricing page, the Studio and platform-overview docs and the API price table; an account was not created and no agent was deployed








