将 Opus 作为顾问，与作为执行者的 Sonnet 或 Haiku 配对，就能以极低的成本让您的智能体获得接近 Opus 级别的智能。
- 分类产品公告
- 产品Claude 平台
- 日期2026 年 4 月 9 日
- 阅读时间5分钟
- https://claude.com/blog/the-advisor-strategy
希望更好地平衡智能与成本的开发者们，已经共同聚焦于我们称之为“顾问策略”的方法：将 Opus 作为顾问，与作为执行者的 Sonnet 或 Haiku 配对。这能让您的智能体获得接近 Opus 级别的智能，同时将成本维持在接近 Sonnet 的水平。
今天，我们在 Claude 平台上推出了顾问工具，让您只需在 API 调用中修改一行代码即可实现顾问策略。
使用顾问策略构建高性价比的智能体
在顾问策略中，Sonnet 或 Haiku 作为执行者端到端地运行任务，调用工具、读取结果并迭代寻求解决方案。当执行者遇到无法合理解决的决策时，它会向作为顾问的 Opus 寻求指导。Opus 访问共享上下文，返回一个计划、修正方案或停止信号，然后执行者继续工作。顾问从不调用工具或产生面向用户的输出，仅向执行者提供指导。
这颠覆了一种常见的子智能体模式，即由一个较大的编排模型分解工作并委派给较小的执行模型。在顾问策略中，一个更小、更具成本效益的模型驱动流程并在必要时升级处理，无需分解、工作池或编排逻辑。前沿级别的推理仅在执行者需要时应用，其余运行过程则保持在执行者级别的成本。
在我们的评估中，使用 Opus 作为顾问的 Sonnet 在 SWE-bench Multilingual1 上的得分比单独使用 Sonnet 提高了 2.7 个百分点，同时每个智能体任务的成本降低了 11.9%。
顾问工具
我们通过顾问工具将顾问策略引入 API，这是一个服务器端工具，Sonnet 和 Haiku 在需要针对特定任务获得指导或帮助时，会知道调用它。
在我们的评估中，配备 Opus 顾问的 Sonnet 在 BrowseComp2 和 Terminal-Bench 2.03 基准测试中的得分均有所提升，同时每个任务的成本低于单独使用 Sonnet。
顾问策略同样适用于 Haiku 作为执行模型。在 BrowseComp 基准测试中，Haiku 搭配 Opus 顾问取得了 41.2% 的分数，是其单独得分 19.7% 的两倍多。Haiku 搭配 Opus 顾问的得分比单独使用 Sonnet 低 29%，但每项任务的成本降低了 85%。与单独使用 Haiku 相比，顾问会增加成本，但组合价格仍仅为 Sonnet 成本的一小部分，使其成为需要兼顾智能与成本的高吞吐量任务的强力选择。
在 Messages API 请求中声明 `advisor_20260301`，模型交接将在单个 `/v1/messages` 请求内完成——无需额外的往返或上下文管理。执行模型决定何时调用它。当调用发生时，我们会将策划好的上下文路由到顾问模型，返回计划，然后执行模型在同一请求内继续执行。
response = client.messages.create(
model="claude-sonnet-4-6", # executor tools=[
{
"type": "advisor_20260301",
"name": "advisor",
"model": "claude-opus-4-6",
"max_uses": 3,
},
# ... your other tools ],
messages=[...]
)
# Advisor tokens reported separately# in the usage block.定价。顾问 token 按顾问模型的费率计费；执行 token 按执行模型的费率计费。由于顾问仅生成简短的计划（通常为 400-700 个文本 token），而执行模型以其较低的费率处理完整输出，因此总体成本远低于端到端运行顾问模型。内置成本控制。设置 `max_uses` 以限制每个请求的顾问调用次数。顾问 token 在使用情况模块中单独报告，以便您按层级跟踪支出。
与您现有工具协同工作。顾问工具只是 Messages API 请求中的另一个条目。您的智能体可以在同一个循环中搜索网络、执行代码以及咨询 Opus。
“它在复杂任务上能做出更好的架构决策，同时在简单任务上不增加任何开销。其制定的计划和轨迹有着天壤之别。”
Eric Simmons，Bolt 首席执行官兼创始人
“我们观察到智能体交互轮次、工具调用次数以及总体得分均有明显提升——优于我们自己构建的规划工具。”
Kay Zhu，Genspark 联合创始人兼首席技术官
“在结构化文档提取任务中，顾问工具使 Haiku 4.5 能够根据复杂度的需求，通过咨询 Opus 4.6 动态扩展智能水平，以 5 倍更低的成本达到前沿模型的质量。”
Anuraj Pandey，Eve Legal 机器学习工程师
`advisor` 工具现已在 Claude 平台上以原生 Beta 版形式提供。要开始使用：
- 添加测试版功能标头：`anthropic-beta: advisor-tool-2026-03-01`
- 在 Messages API 请求中添加 `advisor_20260301` 参数
- 根据您的使用场景修改系统提示词
我们建议您针对以下三种配置运行现有的评估套件：Sonnet 独立运行、Sonnet 执行器搭配 Opus 顾问、以及 Opus 独立运行。查阅文档以了解更多信息。
脚注
- SWE-bench Multilingual：Sonnet 4.6 独立运行时使用了自适应思维。Sonnet 4.6 + Advisor 使用了我们建议的编码系统提示词，并关闭了思维功能。两次运行均采用高努力模式，使用 bash 和文件编辑工具。分数为九种语言共 300 个问题、五次试验的平均值。Opus 4.6 在所有运行中均作为顾问模型使用。
- BrowseComp：所有运行均关闭思维功能，使用网络搜索和网络抓取工具。Sonnet 4.6 运行采用中等努力模式。Sonnet 4.6 + Advisor 使用了我们建议的编码系统提示词；Haiku 4.5 + Advisor 未使用。未使用程序化工具调用或上下文压缩。分数基于 1,266 个问题，每个问题尝试一次。Opus 4.6 在所有运行中均作为顾问模型使用。
- Terminal-Bench 2.0：所有运行均关闭思维功能，使用 bash 和文件编辑工具。Sonnet 4.6 运行采用中等努力模式。两次顾问运行均未使用我们建议的编码系统提示词。每个任务在隔离的 pod 中运行，分配 3 倍资源，超时时间为 1 倍。分数为 89 个任务中每个任务五次尝试的平均值。Opus 4.6 在所有运行中均作为顾问模型使用。
借助 Claude 改变您组织的运作方式
Pair Opus as an advisor with Sonnet or Haiku as an executor, and get near Opus-level intelligence in your agents at a fraction of the cost.
- Category
- ProductClaude Platform
- DateApril 9, 2026
- Reading time5min
- https://claude.com/blog/the-advisor-strategy
Developers who want to better balance intelligence and cost have converged on what we call the advisor strategy: pair Opus as an advisor with Sonnet or Haiku as an executor. This brings near Opus-level intelligence to your agents while keeping costs near Sonnet levels.
Today we're introducing the advisor tool on the Claude Platform to make the advisor strategy a one-line change in your API call.
Build cost-effective agents with the advisor strategy
With the advisor strategy, Sonnet or Haiku runs the task end-to-end as the executor, calling tools, reading results, and iterating toward a solution. When the executor hits a decision it can't reasonably solve, it consults Opus for guidance as the advisor. Opus accesses the shared context and returns a plan, a correction, or a stop signal, and the executor resumes. The advisor never calls tools or produces user-facing output, and only provides guidance to the executor.
This inverts a common sub-agent pattern, where a larger orchestrator model decomposes work and delegates to smaller worker models. In the advisor strategy, a smaller, more cost-effective model drives and escalates without decomposition, a worker pool, or orchestration logic. Frontier-level reasoning applies only when the executor needs it, and the rest of the run stays at executor-level cost.
In our evaluations, Sonnet with Opus as an advisor showed a 2.7 percentage point increase on SWE-bench Multilingual1 over Sonnet alone, while reducing cost per agentic task by 11.9%.
The advisor tool
We’re bringing the advisor strategy to our API with the advisor tool, a server-side tool which Sonnet and Haiku know to invoke when they need guidance or help with a specific task.
In our evaluations, Sonnet with an Opus advisor improved scores across BrowseComp2 and Terminal-Bench 2.03 benchmarks while costing less per task than Sonnet alone.
The advisor strategy also works with Haiku as the executor. On BrowseComp, Haiku with an Opus advisor scored 41.2%, more than double its solo score of 19.7%. Haiku with an Opus advisor trails Sonnet solo by 29% in score but costs 85% less per task. The advisor adds cost relative to Haiku alone, but the combined price is still a fraction of what Sonnet costs, making it a strong option for high-volume tasks that require a balance of intelligence and cost.
Declareadvisor_20260301 in your Messages API request, and the model handoff happens inside a single /v1/messages request—no extra round-trips or context management. The executor model decides when to invoke it. When it does, we route the curated context to the advisor model, return the plan, and the executor continues all within the same request.
response = client.messages.create(
model="claude-sonnet-4-6", # executor tools=[
{
"type": "advisor_20260301",
"name": "advisor",
"model": "claude-opus-4-6",
"max_uses": 3,
},
# ... your other tools ],
messages=[...]
)
# Advisor tokens reported separately# in the usage block.Pricing. Advisor tokens are billed at the advisor model's rates; executor tokens are billed at the executor model's rates. Since the advisor only generates a short plan (typically 400-700 text tokens) while the executor handles the full output at its lower rate, the overall cost stays well below running the advisor model end-to-end.
Built-in cost controls. Set max_uses to cap advisor calls per request. Advisor tokens are reported separately in the usage block so you can track spend per tier.
Works alongside your existing tools. The advisor tool is just another entry in your Messages API request. Your agent can search the web, execute code, and consult Opus in the same loop.
“It makes better architectural decisions on complex tasks while adding no overhead on simple ones. The plans and trajectories are night and day different.”
Eric Simmons, CEO and Founder, Bolt
“We saw clear improvements in agent turns, tool calls, and overall score — better than a planning tool we built ourselves.”
Kay Zhu, Cofounder & CTO, Genspark
“On structured document extraction tasks, the advisor tool enables Haiku 4.5 to dynamically scale intelligence by consulting Opus 4.6 as complexity demands, matching frontier-model quality at 5× lower cost.”
Anuraj Pandey, Machine Learning Engineer, Eve Legal
Theadvisor toolis available now in beta natively on the Claude Platform. To get started:
- Add the beta feature header: anthropic-beta: advisor-tool-2026-03-01
- Add the advisor_20260301 to your Messages API request
- Modify your system prompt based on your use case
We recommend running your existing eval suite against Sonnet solo, Sonnet executor with Opus advisor, and Opus solo. Explore the docs to learn more.
Footnotes
- SWE-bench Multilingual: Sonnet 4.6 solo used adaptive thinking. Sonnet 4.6 + Advisor used our suggested system prompt for coding with thinking turned off. Both runs used high effort with bash and file editing tools. Scores are averaged over five trials of 300 problems across nine languages. Opus 4.6 was used as the advisor model in all runs.
- BrowseComp: All runs used thinking turned off with web search and web fetch tools. Sonnet 4.6 runs used medium effort. Sonnet 4.6 + Advisor used our suggested system prompt for coding; Haiku 4.5 + Advisor did not. No programmatic tool calling or context compaction. Scores are based on 1,266 problems with one attempt per problem. Opus 4.6 was used as the advisor model in all runs.
- Terminal-Bench 2.0: All runs used thinking turned off with bash and file editing tools. Sonnet 4.6 runs used medium effort. Neither advisor run used our suggested system prompt for coding. Each task ran in an isolated pod with 3x resource allocation and a 1x timeout. Scores are averaged over five attempts per task across 89 tasks. Opus 4.6 was used as the advisor model in all runs.