Create Agent Response
Generate a response for the provided input with optional web search and reasoning.
Authorizations
Body
Optional total output-token budget across the run, including for Anthropic models. When neither the request nor its effort level or saved profile supplies a budget, each generation can produce up to 128,000 output tokens, capped by the model's supported limit.
x >= 1Maximum number of research loop steps. If provided, overrides the effort level's max_steps value. For requests that specify model or models without an effort level, the default is 1. Set max_steps to at least 3 when using finance_search with a direct model so the agent has enough steps to initialize and run the tool. Must be >= 1 if specified. Maximum allowed is 500.
1 <= x <= 500Effort to run, as name, name@latest, or name@<version> (e.g., "fast", "low", "fast@2026-09-29.0").
An effort level is a pre-configured model with system prompt and search parameters.
name and name@latest run the active version. name@<version> runs that exact version.
An effort value returned on a response pins that configuration only when it carries a version.
A returned value without a version runs the active version.
Versions are opaque: do not parse, order, or calculate them.
Required if model is not provided.
Stable key for prompt cache affinity. Send the same key on requests that share a prompt prefix, including retries and multi-turn conversations, so later requests can read the cached prefix. The key is scoped to your account and does not affect model routing. See Prompt Caching.
Prompt cache options. implicit (default) keeps the default cache breakpoints and provider automatic caching. explicit turns both off, so only input_text parts marked with prompt_cache_breakpoint are cached; with no marks, the request makes no cache writes. See Prompt Caching.
Tools available to the model
Web search tool configuration for the Responses API
- WebSearchTool
- ImageSearchTool
- FinanceSearchTool
- PeopleSearchTool
- FetchUrlTool
- FunctionTool
- SandboxTool
- McpTool
- ConnectorTool
Built-in, request-scoped inline, and organization-owned custom skills available to the model. Skill metadata is disclosed to the model up front; full instructions are loaded on demand through the load_skill tool. Selecting any skill enables the sandbox tool for the request. Requests with skills run on the durable backend and skills are not echoed back on Response objects.
16- BuiltinSkill
- InlineSkill
- CustomSkill
Response
Successful response. Content type depends on stream parameter:
stream: false(default):application/jsonwith Responsestream: true:text/event-streamwith SSE events
Non-streaming response returned when stream is false
Array of output items (messages, search results, tool calls)
- MessageOutputItem
- SearchResultsOutputItem
- ImageSearchResultsOutputItem
- FetchUrlResultsOutputItem
- FinanceResultsOutputItem
- PeopleSearchResultsOutputItem
- FunctionCallOutputItem
- SandboxResultsOutputItem
- SandboxWriteFileOutputItem
- SandboxReadFileOutputItem
- SkillLoadedOutputItem
- McpListToolsOutputItem
- McpCallOutputItem
- McpApprovalRequestOutputItem
- ToolSearchOutputItem
Effort and version that ran this response, as name@<version> (e.g., "fast@2026-09-29.0").
Present only when the request used an effort level. Streams resumed with starting_after omit it.
For an alias, this is the effort and version that the alias selects.
Send a value with a version as the request effort to run that configuration again.
A released version can receive bug-fix patches, so sending it back runs that version as currently patched.
A value without a version is the bare effort name: the run was created before effort versions, or was served without versions.
Sending it runs the active version.