Skip to main content

Anthropic

This provider supports the Anthropic Claude series of models.

Note: Anthropic models can also be accessed through Azure AI Foundry, AWS Bedrock, and Google Vertex.

Agentic Evals

For agentic evaluations that need built-in file access and skill plugins on top of the Messages API, see the Claude Agent SDK provider. The Messages provider documented here speaks directly to MCP servers via the mcp config so you can plug in your own tools without changing providers.

Setup​

To use Anthropic, you need to set the ANTHROPIC_API_KEY environment variable or specify the apiKey in the provider configuration.

Create Anthropic API keys here.

Example of setting the environment variable:

export ANTHROPIC_API_KEY=your_api_key_here

Authenticating via a Claude Code session​

If you already have an active Claude Code session (for example as a Claude Pro or Max subscriber), you can reuse its OAuth credential instead of creating a separate Anthropic Console API key. Set apiKeyRequired: false on the provider config:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      apiKeyRequired: false

When apiKeyRequired is false and no ANTHROPIC_API_KEY is available, Promptfoo loads the Claude Code OAuth credential from:

  1. The macOS keychain entry Claude Code-credentials (darwin only), then
  2. $HOME/.claude/.credentials.json on Linux and macOS, or %USERPROFILE%\.claude\.credentials.json on Windows.

Set CLAUDE_CONFIG_DIR to read the credential from a different Claude Code profile — the same environment variable the Claude Code CLI itself uses to relocate ~/.claude. It can be set in your shell, in the config's top-level env: block, or in a provider's env: block (the provider-scoped value wins). On macOS, where Claude Code stores credentials in the system keychain, Promptfoo mirrors the CLI's profile-specific keychain entry: when CLAUDE_CONFIG_DIR is set, the credential is looked up under that profile's keychain service (derived from the configured directory) rather than the default one, so evals authenticate as the profile you selected.

Promptfoo authenticates requests with a Bearer token, sends the claude-code-20250219,oauth-2025-04-20 beta headers, and prepends the required Claude Code identity system block ("You are Claude Code, Anthropic's official CLI for Claude.") to every Messages request. Your own system prompt is still forwarded as the next system block.

If you haven't logged in yet, run claude /login to create a credential. Re-run it if Promptfoo warns that the credential has expired. Requests made this way are expected to count against your Claude subscription the same way calls from the Claude Code CLI do — check Anthropic's documentation for current billing behavior.

This also enables model-graded assertions such as llm-rubric to run without a separate Anthropic Console key — see the example below.

Models​

These models currently resolve on the Anthropic Messages API:

Model IDDescription
anthropic:messages:claude-fable-5-1Claude Fable 5.1
anthropic:messages:claude-mythos-5-1Claude Mythos 5.1
anthropic:messages:claude-fable-5Claude Fable 5
anthropic:messages:claude-mythos-5Claude Mythos 5
anthropic:messages:claude-opus-5-5Claude Opus 5.5
anthropic:messages:claude-sonnet-5-5Claude Sonnet 5.5
anthropic:messages:claude-opus-5Claude Opus 5
anthropic:messages:claude-opus-4-8Claude 4.8 Opus
anthropic:messages:claude-opus-4-7Claude 4.7 Opus
anthropic:messages:claude-sonnet-5Claude Sonnet 5
anthropic:messages:claude-sonnet-4-6Claude 4.6 Sonnet
anthropic:messages:claude-opus-4-6Claude 4.6 Opus
anthropic:messages:claude-opus-4-5-20251101 (claude-opus-4-5)Claude 4.5 Opus
anthropic:messages:claude-sonnet-4-5-20250929 (claude-sonnet-4-5)Claude 4.5 Sonnet
anthropic:messages:claude-haiku-4-5-20251001 (claude-haiku-4-5)Claude 4.5 Haiku

The Mythos rows are limited-access models: the ID is correct, but an org without access gets the same not_found_error a retired model returns.

Retired on the Anthropic API​

These IDs return 404 not_found_error from Anthropic, so a direct anthropic:messages: call will fail. Promptfoo still keeps their pricing, because cost attribution on historical evals needs it and because partner platforms set their own lifecycle dates.

Availability elsewhere is per-model, not a blanket rule — check the row in Cross-Platform Model Availability before assuming a retired ID still works somewhere. Three of these are withdrawn from Bedrock as well (claude-3-opus-20240229, claude-opus-4-20250514, claude-3-5-haiku-20241022) and are rejected locally with Unknown Amazon Bedrock model.

Model IDDescriptionSuggested replacement
claude-opus-4-1-20250805Claude 4.1 Opusclaude-opus-5-5
claude-opus-4-20250514Claude 4 Opusclaude-opus-5-5
claude-sonnet-4-20250514Claude 4 Sonnetclaude-sonnet-5
claude-3-7-sonnet-20250219Claude 3.7 Sonnetclaude-sonnet-5
claude-3-5-sonnet-20241022Claude 3.5 Sonnet (v2)claude-sonnet-5
claude-3-5-sonnet-20240620Claude 3.5 Sonnet (v1)claude-sonnet-5
claude-3-5-haiku-20241022Claude 3.5 Haikuclaude-haiku-4-5
claude-3-opus-20240229Claude 3 Opusclaude-opus-5-5
claude-3-haiku-20240307Claude 3 Haikuclaude-haiku-4-5
Model aliases

Anthropic does not publish -latest aliases — claude-sonnet-4-5-latest returns a not_found_error. Where a shorter alias exists it is the bare family ID shown in parentheses above: claude-sonnet-4-5 resolves to claude-sonnet-4-5-20250929. Claude 4.6 and newer are already unversioned IDs, so there is nothing to shorten.

Rows without a parenthetical have no alias. Retired on the Anthropic API lists retirements from the direct Anthropic API. Check the platform guides below for availability through other providers.

Cross-Platform Model Availability​

Claude models are available across multiple platforms. Here's how the model names map across different providers:

ModelAnthropic APIAzure AI Foundry (docs)AWS Bedrock (docs)GCP Vertex AI (docs)
Claude Fable 5.1claude-fable-5-1claude-fable-5-1anthropic.claude-fable-5-1claude-fable-5-1
Claude Mythos 5.1claude-mythos-5-1claude-mythos-5-1 (limited)anthropic.claude-mythos-5-1 (limited)claude-mythos-5-1 (limited)
Claude Fable 5claude-fable-5claude-fable-5anthropic.claude-fable-5claude-fable-5
Claude Mythos 5claude-mythos-5Not availableanthropic.claude-mythos-5 (limited)Limited availability; ID not public
Claude Opus 5.5claude-opus-5-5claude-opus-5-5anthropic.claude-opus-5-5claude-opus-5-5
Claude Sonnet 5.5claude-sonnet-5-5claude-sonnet-5-5global.anthropic.claude-sonnet-5-5claude-sonnet-5-5
Claude Opus 5claude-opus-5claude-opus-5anthropic.claude-opus-5claude-opus-5
Claude 4.8 Opusclaude-opus-4-8claude-opus-4-8anthropic.claude-opus-4-8claude-opus-4-8
Claude 4.7 Opusclaude-opus-4-7claude-opus-4-7anthropic.claude-opus-4-7claude-opus-4-7
Claude Sonnet 5claude-sonnet-5claude-sonnet-5anthropic.claude-sonnet-5claude-sonnet-5
Claude 4.6 Sonnetclaude-sonnet-4-6claude-sonnet-4-6anthropic.claude-sonnet-4-6claude-sonnet-4-6
Claude 4.6 Opusclaude-opus-4-6claude-opus-4-6-20260205anthropic.claude-opus-4-6-v1claude-opus-4-6
Claude 4.5 Opusclaude-opus-4-5-20251101 (claude-opus-4-5)claude-opus-4-5-20251101anthropic.claude-opus-4-5-20251101-v1:0claude-opus-4-5@20251101
Claude 4.5 Sonnetclaude-sonnet-4-5-20250929 (claude-sonnet-4-5)claude-sonnet-4-5-20250929anthropic.claude-sonnet-4-5-20250929-v1:0claude-sonnet-4-5@20250929
Claude 4.5 Haikuclaude-haiku-4-5-20251001 (claude-haiku-4-5)claude-haiku-4-5-20251001anthropic.claude-haiku-4-5-20251001-v1:0claude-haiku-4-5@20251001
Claude 4.1 OpusRetired on the direct APIclaude-opus-4-1-20250805anthropic.claude-opus-4-1-20250805-v1:0claude-opus-4-1@20250805
Claude 4 OpusRetired on the direct APIclaude-opus-4-20250514Withdrawn from Bedrockclaude-opus-4@20250514
Claude 4 SonnetRetired on the direct APIclaude-sonnet-4-20250514anthropic.claude-sonnet-4-20250514-v1:0claude-sonnet-4@20250514
Claude 3.7 SonnetRetired on the direct APIclaude-3-7-sonnet-20250219anthropic.claude-3-7-sonnet-20250219-v1:0claude-3-7-sonnet@20250219
Claude 3.5 SonnetRetired on the direct APIclaude-3-5-sonnet-20241022anthropic.claude-3-5-sonnet-20241022-v2:0claude-3-5-sonnet-v2@20241022
Claude 3.5 HaikuRetired on the direct APIclaude-3-5-haiku-20241022Withdrawn from Bedrockclaude-3-5-haiku@20241022
Claude 3 OpusRetired on the direct APIclaude-3-opus-20240229Withdrawn from Bedrockclaude-3-opus@20240229
Claude 3 HaikuRetired on the direct APIclaude-3-haiku-20240307anthropic.claude-3-haiku-20240307-v1:0claude-3-haiku@20240307

Supported Parameters​

Config PropertyEnvironment VariableDescription
apiKeyANTHROPIC_API_KEYYour API key from Anthropic
apiKeyRequired-Skip the API key preflight and authenticate via a local Claude Code session
apiBaseUrlANTHROPIC_BASE_URLThe base URL for requests to the Anthropic API
temperatureANTHROPIC_TEMPERATUREControls the randomness of the output (default: 0). Omitted when top_p is set.
max_tokensANTHROPIC_MAX_TOKENSThe maximum length of the generated text (default: 1024, or 2048 whenever thinking will consume output tokens — either because you enabled it or because the model thinks by default)
cost-Legacy per-token override applied to both input and output pricing
inputCost-Override input token pricing in promptfoo cost estimates
outputCost-Override output token pricing in promptfoo cost estimates
top_p-Controls nucleus sampling. Mutually exclusive with temperature.
top_k-Only sample from the top K options for each subsequent token
stop_sequences-Array of strings that will stop generation when encountered
stream-Enable streaming (required when max_tokens > 21,333)
tools-An array of tool or function definitions for the model to call
tool_choice-An object specifying the tool to call
effort-Output effort level: low, medium, high, xhigh, or max
output_format-JSON schema configuration for structured outputs
thinking-Configuration for Claude's extended thinking (enabled, adaptive, disabled, or between_tools)
showThinking-Whether to include thinking content in the output (default: true)
cache_control-Auto-apply cache_control to the last cacheable block in the request
metadata-Request metadata such as user_id for tracking purposes
service_tier-Priority tier: auto (default) or standard_only
beta-Array of anthropic-beta feature flags to send with the request
headers-Additional headers to be sent with the API request
extra_body-Additional parameters to be included in the API request body

For MCP tools, set mcp.enabled: true. The max_tool_calls option caps MCP tool executions per request (default: 8).

Sampling support varies by model. Promptfoo omits unsupported temperature, top_p, and top_k settings; see the model notes below.

Prompt Template​

To allow for compatibility with the OpenAI prompt template, the following format is supported:

prompt.json
[
  {
    "role": "system",
    "content": "{{ system_message }}"
  },
  {
    "role": "user",
    "content": "{{ question }}"
  }
]

Promptfoo extracts system messages into the API's system prompt and forwards user and assistant messages. Set system_message and question in your test's vars.

Assistant prefill

Claude 4.6 and later models reject prompts ending with an assistant message. End with a user message instead. To constrain the response format, use structured outputs or a system instruction. The 4.5 models still accept assistant prefill.

Options​

Set supported parameters under config:

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      effort: medium
      max_tokens: 2048 # Includes thinking and the final answer
prompts:
  - file://prompt.json

Stop Sequences​

Use stop_sequences to halt generation when Claude encounters specific strings:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      stop_sequences:
        - "\n\nHuman:"
        - 'STOP'

Classifier Refusals​

An Anthropic safety-classifier refusal can arrive as a successful Messages response with stop_reason: refusal. When the response also includes stop_details, Promptfoo returns the provider content as output and normalizes:

{
  "finishReason": "content_filter",
  "guardrails": {
    "flagged": true,
    "reason": "Content refused by Anthropic safety filters — category: ..."
  }
}

Promptfoo currently requires stop_details to create the top-level guardrails signal. A model-written refusal or stop_reason: refusal without details is therefore not the same assertion result. Treat the optional stop_details.category and stop_details.explanation as diagnostic evidence. API validation errors remain provider errors and skip assertions.

In a stream, the terminal refusal reason can arrive after partial text in the final message delta. Promptfoo merges those details before returning the provider response, and cached structured refusals preserve the signal. Use not-guardrails to require the structured classifier signal. Use is-refusal for model-written refusal text.

Cost estimates follow Anthropic's refusal billing rules: refusals before any output cost zero unless the category is bio, frontier_llm, or reasoning_extraction. Those categories and refusals after output begins use normal token pricing.

Metadata​

Pass request metadata to the API for tracking or auditing purposes:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      metadata:
        user_id: 'user-123'

Tool Calling​

The Anthropic provider supports tool calling (function calling). Here's an example configuration for defining tools.

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      tools:
        - name: get_weather
          description: Get the current weather in a given location
          input_schema:
            type: object
            properties:
              location:
                type: string
                description: The city and state, e.g., San Francisco, CA
              unit:
                type: string
                enum:
                  - celsius
                  - fahrenheit
            required:
              - location

Web Search and Web Fetch Tools​

Anthropic provides specialized tools for web search and web fetching capabilities:

Web Fetch Tool​

The web fetch tool allows Claude to retrieve full content from web pages and PDF documents. This is useful when you want Claude to access and analyze specific web content.

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      tools:
        - type: web_fetch_20250910
          name: web_fetch
          max_uses: 5
          allowed_domains:
            - docs.example.com
            - help.example.com
          citations:
            enabled: true
          max_content_tokens: 50000

Use one fetch version per request. web_fetch_20260209 adds dynamic filtering, web_fetch_20260309 adds cache control, and web_fetch_20260318 adds response inclusion control:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      tools:
        - type: web_fetch_20260318
          name: web_fetch
          max_uses: 3
          use_cache: false
          response_inclusion: excluded

response_inclusion: excluded omits nested tool-use/result pairs consumed by a completed code-execution call. Direct fetches and paused calls still return their full results. See Anthropic's web fetch documentation.

Web Fetch Tool Configuration Options:

ParameterTypeDescription
typestringweb_fetch_20250910 (beta), web_fetch_20260209, web_fetch_20260309 (adds use_cache), or web_fetch_20260318 (adds response_inclusion)
namestringMust be web_fetch
max_usesnumberMaximum number of web fetches per request (optional)
allowed_callersstring[]Restrict which tool callers may invoke the server tool (optional)
allowed_domainsstring[]List of domains to allow fetching from (optional, mutually exclusive with blocked_domains)
blocked_domainsstring[]List of domains to block fetching from (optional, mutually exclusive with allowed_domains)
defer_loadingbooleanLoad the tool lazily instead of including it in the initial system prompt (optional)
citationsobjectEnable citations with { enabled: true } (optional)
max_content_tokensnumberMaximum tokens for web content (optional)
cache_controlobjectApply Anthropic cache control to the tool definition (optional)
strictbooleanEnable strict schema validation for tool names and inputs (optional)
url_sourcesobjectLimit fetchable URLs by source: user_input, client_tool_results, or server_tool_results, each with a tagged filter such as { type: none } (optional)
use_cachebooleanWhether to use cached content (web_fetch_20260309 and web_fetch_20260318, optional)
response_inclusionstringfull (default) or excluded; applies to completed nested code-execution calls (web_fetch_20260318 only)
Web Search Tool​

The web search tool allows Claude to search the internet for information:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      tools:
        - type: web_search_20260209
          name: web_search
          max_uses: 3

Web Search Tool Configuration Options:

ParameterTypeDescription
typestringweb_search_20250305, web_search_20260209, or web_search_20260318 (adds response_inclusion)
namestringMust be web_search
max_usesnumberMaximum number of searches per request (optional)
allowed_callersstring[]Restrict which tool callers may invoke the server tool (optional)
allowed_domainsstring[]Restrict results to specific domains (optional, mutually exclusive with blocked_domains)
blocked_domainsstring[]Exclude domains from results (optional, mutually exclusive with allowed_domains)
cache_controlobjectApply Anthropic cache control to the tool definition (optional)
defer_loadingbooleanLoad the tool lazily instead of including it in the initial system prompt (optional)
strictbooleanEnable strict schema validation for tool names and inputs (optional)
response_inclusionstringfull (default) or excluded — see the web fetch table above (web_search_20260318 only, optional)
user_locationobjectApproximate user location to improve search relevance (optional)
Combined Web Search and Web Fetch​

Use web search to find URLs and web fetch to read their contents:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      tools:
        - type: web_search_20260209
          name: web_search
          max_uses: 3
        - type: web_fetch_20260309
          name: web_fetch
          max_uses: 5
          citations:
            enabled: true

This configuration allows the model to first search for relevant information, then fetch full content from the most promising results.

Paused Turns​

A long server-tool run can stop with stop_reason: pause_turn before Claude finishes. Promptfoo resumes it up to 5 times, preserving the code execution container and summing each request's usage and cost. Structured output uses the final response's JSON; generated file references remain in metadata.fileReferences. If resuming fails or reaches the limit, the result keeps its partial output with finishReason: pause_turn, logs a warning, and is not cached. Cancellation stops further requests.

Memory Tool​

Anthropic's memory_20250818 tool can be included in tools. Promptfoo passes this native tool definition through unchanged, which is useful for evaluating whether a model requests memory operations. Promptfoo does not manage Anthropic memory stores or run local memory handlers for you.

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      tools:
        - type: memory_20250818
          name: memory
          allowed_callers:
            - direct

Memory Tool Configuration Options:

ParameterTypeDescription
typestringMust be memory_20250818
namestringMust be memory
allowed_callersstring[]Restrict which tool callers may invoke the memory tool (optional)
cache_controlobjectApply Anthropic cache control to the tool definition (optional)
defer_loadingbooleanLoad the tool lazily instead of including it in the initial prompt
input_examplesobject[]Example memory commands to include in the tool definition (optional)
strictbooleanEnable strict schema validation for tool names and inputs (optional)

Important Security Notes:

  • The web fetch tool requires trusted environments due to potential data exfiltration risks
  • The model cannot dynamically construct URLs - only URLs provided by users or from search results can be fetched
  • Use domain filtering to restrict access to specific sites:
    • Use allowed_domains to whitelist trusted domains (recommended)
    • Use blocked_domains to blacklist specific domains
    • Note: Only one of allowed_domains or blocked_domains can be specified, not both

Model Context Protocol (MCP)​

The Anthropic Messages provider can connect to any MCP server — stdio, SSE, or streamable HTTP — and execute the model's tool_use blocks against that server, feeding the tool_result back into the conversation until Claude produces a final reply.

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      mcp:
        enabled: true
        # Inline command-based stdio server, or `path` to a local script
        server:
          command: npx
          args: ['-y', '@modelcontextprotocol/server-filesystem', '/tmp/workspace']
        # Or use a remote SSE / streamable HTTP server
        # servers:
        #   - name: deepwiki
        #     url: https://mcp.deepwiki.com/mcp
      # Optional cap on MCP rounds per request (default 8). Enforced locally;
      # not sent to Anthropic.
      max_tool_calls: 5

How it works:

  • Tools discovered on the MCP server are passed to Claude alongside any inline tools.
  • When Claude returns a tool_use block whose name matches an MCP tool, promptfoo calls the tool with the model's arguments and appends a matching tool_result block on the user turn.
  • The loop repeats until Claude returns text (no more tool_use) or max_tool_calls is hit. Tool errors are forwarded as tool_result blocks with is_error: true so the model can recover.
  • Non-MCP tool_use blocks (regular function tools, or built-ins like web_search) are passed through to the existing output and not auto-executed.
Response caching with MCP

The disk response cache is skipped while mcp.enabled is true, because tool results can be non-deterministic between runs. Use max_tool_calls to bound spend.

See the MCP integration guide for full server configuration options (auth, timeouts, multiple servers, etc.) and the Anthropic MCP example.

See the Anthropic Tool Use Guide for more information on how to define tools and the tool use example here.

Images / Vision​

All current Claude models accept images in the prompt.

See the Claude vision example.

Claude accepts base64 images and image URLs. Use Anthropic image content blocks; their shape differs from OpenAI's image_url blocks. See the vision API guide for request examples.

Prompt Caching​

Prompt caching reuses unchanged prompt prefixes and is supported by current Claude models. Mark the prefix to cache with cache_control:

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-sonnet-5
prompts:
  - file://prompts.yaml
prompts.yaml
- role: system
  content:
    - type: text
      text: 'System message'
      cache_control:
        type: ephemeral
    - type: text
      text: '{{context}}'
      cache_control:
        type: ephemeral
- role: user
  content: '{{question}}'

As a simpler alternative, use the top-level cache_control parameter to automatically apply a cache marker to the last cacheable block in the request, without annotating each block individually:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      cache_control:
        type: ephemeral

Common use cases for caching:

  • System messages and instructions
  • Tool/function definitions
  • Large context documents
  • Frequently used images

Cache read and creation token counts are tracked in the response's token usage details. Cost estimates use the reported cache duration: five-minute writes cost 1.25 times the base input rate, and one-hour writes cost twice the base input rate.

See Anthropic's Prompt Caching Guide for more details on requirements, pricing, and best practices.

Citations​

Claude can provide detailed citations when answering questions about documents. Basic example:

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-sonnet-5
prompts:
  - file://prompts.yaml
prompts.yaml
- role: user
  content:
    - type: document
      source:
        type: text
        media_type: text/plain
        data: 'Your document text here'
      citations:
        enabled: true
    - type: text
      text: 'Your question here'

See Anthropic's Citations Guide for more details.

PDF Documents​

Claude can process PDF files using document content blocks. Pass the PDF as base64-encoded data:

- role: user
  content:
    - type: document
      source:
        type: base64
        media_type: application/pdf
        data: '{{pdf_base64}}'
    - type: text
      text: 'Summarize this document'

Use a test var to supply the base64-encoded PDF content:

tests:
  - vars:
      pdf_base64: file://document.pdf

Claude Fable 5.1 and Mythos 5.1​

Use the pinned model IDs below. Mythos 5.1 requires Project Glasswing access and provider approval.

providers:
  - id: anthropic:messages:claude-fable-5-1
    config:
      max_tokens: 4096
      effort: high
  - id: anthropic:messages:claude-mythos-5-1
    config:
      max_tokens: 4096
      effort: high

Both models have a 1M-token context window and support up to 128K output tokens. Input and output cost $10 and $50 per million tokens, respectively. Cache reads cost $0.25 per million tokens, down from $1 on Fable 5 and Mythos 5; promptfoo includes this discount in its cost estimates. See Anthropic's pricing.

Thinking is always on, and the sampling and thinking normalization described below also applies to 5.1. Unlike Fable 5, 5.1 rejects forced tool use: promptfoo omits tool_choice with type any or tool and warns. Use auto or none instead. When replaying Fable 5.1 thinking blocks, keep earlier messages, system prompts, and tools unchanged; edited prefixes can cause API errors. See Anthropic's migration notes.

Claude Fable 5 and Mythos 5 notes​

Fable 5 and Mythos 5 use always-on adaptive thinking. Promptfoo omits unsupported temperature, top_p, and top_k values, converts legacy thinking: { type: 'enabled', budget_tokens: N } configs to adaptive thinking, and omits thinking: { type: 'disabled' } because thinking cannot be disabled. Set thinking: { type: 'adaptive', display: 'summarized' } to include a readable thinking summary; the default display: 'omitted' returns an empty thinking block, which Promptfoo excludes from the output.

Both models use a 1M-token context window, support up to 128K output tokens, and are priced at $10 per million input tokens and $50 per million output tokens. Mythos 5 access is limited through Project Glasswing and may require provider approval. Both model IDs are pinned.

Claude Opus 5.5 notes​

Opus 5.5 is priced below Opus 5 and has the same context window, maximum output, and tokenizer. Its request rules match Fable 5.1's, and promptfoo adjusts requests to follow them:

  • Thinking is always on. Opus 5.5 rejects thinking: { type: 'disabled' } at every effort level, so promptfoo removes it and logs a warning once. Manual thinking: { type: 'enabled', budget_tokens: N } configs become thinking: { type: 'adaptive' }.
  • effort defaults to medium, one level below Opus 5's high. Set effort explicitly when you compare the two. It is the only way to control how much the model thinks.
  • Forced tool use and sampling controls are rejected. Promptfoo omits tool_choice values of type any or tool (use auto or none) and all of temperature, top_p, and top_k.

Opus 5.5 costs a flat $4 per million input tokens and $20 per million output tokens across its 1M-token context window. Cache reads cost $0.20 per million tokens, and promptfoo's cost estimates include that rate. To track Anthropic's fast mode, set inputCost: 8 / 1e6 and outputCost: 40 / 1e6.

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-opus-5-5
    config:
      effort: high
      max_tokens: 16000

Claude Opus 5 notes​

Opus 5 is the Opus-tier Claude 5 model, aimed at complex agentic coding and long-horizon work. It keeps Opus 4.8's request surface and pricing, with two behavior changes promptfoo handles for you:

  • Thinking is on by default. Unlike Opus 4.7/4.8 — where omitting thinking meant no extended thinking — an omitted thinking block on Opus 5 runs adaptive thinking. Because max_tokens caps thinking plus response text, promptfoo sizes its default max_tokens with thinking headroom (2048 instead of 1024) so responses aren't truncated mid-answer. Set max_tokens explicitly for anything longer.
  • Disabling thinking is effort-gated. thinking: { type: 'disabled' } is only accepted at effort high or below; pairing it with xhigh or max returns a 400. Promptfoo drops the rejected thinking: { type: 'disabled' } (keeping your effort) and logs a one-time warning. Lower effort to high if you actually need thinking off.
  • Sampling controls are managed for you. Like Opus 4.7/4.8, Opus 5 rejects temperature, top_p, and top_k with a 400; promptfoo omits all three from every request, including its built-in temperature: 0 default. A legacy thinking: { type: 'enabled', budget_tokens: N } config is converted to thinking: { type: 'adaptive' }.
  • The full low → max effort ladder is available. Start at xhigh for coding and agentic work, then sweep downward — low and medium are unusually strong on this model and are the main cost and latency lever. See the Effort Level section.

Opus 5 uses a 1M-token context window (both the default and the maximum) billed at a flat $5 per million input / $25 per million output — the same list rates as Opus 4.8, with no long-context surcharge above 200K tokens. Anthropic's fast mode ($10 / $50, Claude API only) is a separate research-preview rate that promptfoo does not encode. To track it, set inputCost: 10 / 1e6 and outputCost: 50 / 1e6 — a single cost cannot express asymmetric rates, because it is applied as both the input and the output per-token price.

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-opus-5
    config:
      effort: xhigh
      max_tokens: 8192

Claude Sonnet 5.5 notes​

Sonnet 5.5 has a 1M-token context window and a 128K-token output limit. Promptfoo adjusts requests to its API requirements:

  • Thinking is on by default, and disabled is rejected. Sonnet 5.5's lowest thinking setting is thinking: { type: 'between_tools' }, which turns off up-front thinking. Promptfoo sends between_tools in place of disabled and logs a warning once. between_tools is only accepted at effort high or below, so with xhigh or max promptfoo omits disabled or between_tools, warns, and the model thinks adaptively.
  • effort defaults to high. Compare effort levels for your workload when migrating.
  • Forced tool use is rejected. Promptfoo omits tool_choice values of type any or tool. Use auto or none, and say in the prompt when a tool applies.
  • Sampling controls and manual budgets are rejected, as on Sonnet 5. Promptfoo omits temperature, top_p, and top_k, and converts thinking: { type: 'enabled', budget_tokens: N } to thinking: { type: 'adaptive' }.
  • Text between tool calls comes back in thinking blocks. With the default display: 'omitted' those blocks are empty. Set thinking: { type: 'adaptive', display: 'summarized' } or use between_tools to keep that text in the output.

Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens across its 1M-token context window, with cache reads at $0.20 per million tokens.

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-sonnet-5-5
    config:
      effort: medium
      thinking:
        type: between_tools # no up-front thinking; accepted at effort high or below
      max_tokens: 4096

Claude Sonnet 5 notes​

Sonnet 5 has a 1M-token context window and supports effort levels from low through max. It does not support manual sampling controls:

  • Sampling controls: Promptfoo omits temperature, top_p, and top_k, including its default temperature. Explicit settings trigger a one-time warning. The parameters are also omitted through AWS Bedrock, GCP Vertex, and Azure AI Foundry.

  • Manual thinking budgets convert to adaptive. A legacy thinking: { type: 'enabled', budget_tokens: N } config is converted to thinking: { type: 'adaptive' }; use effort to control reasoning depth.

  • Adaptive thinking is on by default. Requests without a thinking field still use adaptive thinking. Pass thinking: { type: 'disabled' } to turn it off, and leave enough max_tokens headroom for thinking plus the visible response.

Sonnet 5 uses a 1M-token context window billed at $2 per million input / $10 per million output, with no long-context surcharge above 200K tokens. Anthropic made these rates permanent on August 10, 2026, canceling the planned September increase. See Anthropic's pricing documentation. The newer tokenizer can produce more tokens for the same text, so compare total request costs when migrating from Sonnet 4.6.

Claude Opus 4.8 notes​

Opus 4.8 uses the same sampling and thinking settings as Opus 4.7:

  • Sampling controls are managed for you. Like Opus 4.7, Opus 4.8 samples adaptively and rejects temperature, top_p, and top_k (any of them returns a 400); promptfoo omits all three from every request. Setting any of them in config or ANTHROPIC_TEMPERATURE logs a one-time heads-up so you can clean the values out of your eval.
  • Adaptive thinking is opt-in. Set thinking: { type: 'adaptive' } to let the model decide how much to reason per request. Without an explicit thinking block the model runs without extended thinking, even at high effort. Manual budget-based thinking (thinking: { type: 'enabled', budget_tokens: N }) is rejected with a 400.
  • effort defaults to high and xhigh is available. Setting effort: high behaves the same as omitting it. Start with xhigh for coding and agentic work. See the Effort Level section.

The same suppression applies when you reach Opus 4.8 through AWS Bedrock, GCP Vertex, or Azure AI Foundry — promptfoo omits the unsupported sampling parameters on each of those paths too (silently; the one-time warning above is specific to the Anthropic Messages provider).

Claude Opus 4.7 notes​

Opus 4.7 supports adaptive thinking when explicitly enabled:

  • Temperature is managed for you. Opus 4.7 samples adaptively and does not accept temperature; promptfoo omits the field from every request. Passing temperature in config or ANTHROPIC_TEMPERATURE logs a one-time heads-up so you can clean the value out of your eval.
  • Adaptive thinking is opt-in. Set thinking: { type: 'adaptive' } to let the model choose how much to reason per request; leaving thinking unset runs Opus 4.7 without extended thinking, even at high effort. (Opus 5 is the model where an omitted block means adaptive.) Budget-based modes from older models aren't used on 4.7.
  • xhigh effort level is available. It sits between high and max and is a good starting point for coding and agentic tasks. See the Effort Level section.
  • Updated tokenizer. The same input can map to 1.0–1.35× more tokens than Opus 4.6, so measure real traffic if you're comparing costs.

The same guidance applies when you reach Opus 4.7 through AWS Bedrock, GCP Vertex, or Azure AI Foundry — promptfoo suppresses temperature on each of those paths as well.

Extended Thinking​

Use thinking to configure reasoning before the final answer. When available, display: summarized returns a summary of that reasoning:

promptfooconfig.yaml
providers:
  # Adaptive thinking
  - id: anthropic:messages:claude-opus-5
    config:
      max_tokens: 20000
      thinking:
        type: 'adaptive'
        display: 'summarized' # Opt in to readable reasoning; default is 'omitted'
      effort: xhigh # Controls reasoning depth instead of a token budget

  # Manual thinking budget — Opus 4.6 / Sonnet 4.6 and older (deprecated on 4.6)
  - id: anthropic:messages:claude-haiku-4-5
    config:
      max_tokens: 20000
      thinking:
        type: 'enabled'
        budget_tokens: 16000 # Must be ≥1024 and less than max_tokens

The thinking configuration has three possible values:

  1. Adaptive thinking (supported on Claude 4.6 and later):
thinking:
  type: 'adaptive'

In adaptive mode, Claude decides when and how much to think based on the complexity of the request. Control depth with effort rather than a token budget.

  1. Manual thinking budgets (Claude 4.5 and 4.6):
thinking:
  type: 'enabled'
  budget_tokens: 16000 # Must be at least 1024 and less than max_tokens

Claude 4.5 and 4.6 accept manual budgets; this mode is deprecated on 4.6. On adaptive-only models, Promptfoo converts enabled to adaptive, removes the budget, and logs a warning.

  1. Disabled thinking:
thinking:
  type: 'disabled'

Not accepted on Opus 5.5 or Fable 5 / Mythos 5 / Fable 5.1 / Mythos 5.1, where thinking is always on, nor on Opus 5 above effort: high. Promptfoo omits it in both cases and warns. On Sonnet 5.5, it becomes between_tools at high effort or below and is omitted at higher effort.

  1. Between tools (Claude Sonnet 5.5 only):
thinking:
  type: between_tools

This mode turns off up-front thinking and requires effort: high or below. Progress updates between tool calls still return as thinking blocks. It takes no other fields, including display or budget_tokens. At xhigh or max, Promptfoo omits it so the model uses adaptive thinking.

Thinking display​

display controls the reasoning text returned by the API. Omitting that text does not disable thinking or remove its token cost:

ValueBehavior
'omitted'Default on Fable 5/5.1, Mythos 5/5.1, Opus 5.5, Opus 5, Opus 4.7/4.8, and Sonnet 5. The thinking block is returned with empty text plus a signature for multi-turn continuity.
'summarized'Returns a readable summary of the reasoning. Default on Claude 4.6 and earlier.

Set display: summarized to request a readable summary:

thinking:
  type: adaptive
  display: summarized

Thinking tokens count toward max_tokens in both modes. Only manual thinking requires budget_tokens, which must be at least 1,024 and below max_tokens.

Promptfoo omits temperature and top_k when thinking is enabled, and clamps top_p to [0.95, 1.0] on models that support it. On models without sampling controls, all three parameters are omitted.

Forced tool use (tool_choice type any or tool) is incompatible with manual thinking and with Opus 5.5, Sonnet 5.5, Fable 5.1, and Mythos 5.1. Promptfoo omits it with a warning in those cases; use auto or none. Other adaptive models accept forced tool use.

Example response with thinking enabled:

{
  "content": [
    {
      "type": "thinking",
      "thinking": "Let me analyze this step by step...",
      "signature": "WaUjzkypQ2mUEVM36O2TxuC06KN8xyfbJwyem2dw3URve/op91XWHOEBLLqIOMfFG/UvLEczmEsUjavL...."
    },
    {
      "type": "text",
      "text": "Based on my analysis, here is the answer..."
    }
  ]
}

Controlling Thinking Output​

thinking.display controls what the API returns; showThinking controls whether Promptfoo includes that text in the output (default: true). It cannot reveal reasoning omitted by the API. For example, request a summary but exclude it from the output being graded:

promptfooconfig.yaml
providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      thinking:
        type: 'adaptive'
        display: 'summarized'
      showThinking: false # Exclude thinking content from the output

Redacted Thinking​

Sometimes Claude's internal reasoning may be flagged by safety systems. When this occurs, the thinking block will be encrypted and returned as a redacted_thinking block:

{
  "content": [
    {
      "type": "redacted_thinking",
      "data": "EmwKAhgBEgy3va3pzix/LafPsn4aDFIT2Xlxh0L5L8rLVyIwxtE3rAFBa8cr3qpP..."
    },
    {
      "type": "text",
      "text": "Based on my analysis..."
    }
  ]
}

Redacted thinking blocks are automatically decrypted when passed back to the API, allowing Claude to maintain context without compromising safety guardrails.

Extended Output with Thinking​

For longer responses, raise max_tokens and enable streaming:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      max_tokens: 64000 # Sonnet 5 supports up to 128K output tokens
      stream: true # Required above 21,333 output tokens
      thinking:
        type: 'adaptive'
      effort: high

Fable 5/5.1, Opus 5.5, Opus 5, Opus 4.6–4.8, Sonnet 5, and Sonnet 4.6 all support up to 128K output tokens; the 4.5 generation caps at 64K. The output-128k-2025-02-19 beta feature is specific to Claude 3.7 Sonnet and is not needed on any current model.

When using extended output:

  • Streaming is required when max_tokens is greater than 21,333
  • Thinking shares the max_tokens budget with the answer, so leave headroom for both
  • The model may not use the entire allocated budget

See Anthropic's Extended Thinking Guide for more details on requirements and best practices.

Effort Level​

effort controls how many tokens Claude spends on a response, including thinking and tool calls. Higher settings can increase quality, cost, and latency. It is not a sampling-temperature control or a hard token limit.

providers:
  - id: anthropic:messages:claude-opus-5
    config:
      effort: xhigh # Options: low, medium, high, xhigh, max

Support varies by model — sending an unsupported level returns a 400:

ModelSupported levels
Fable/Mythos 5 and 5.1, Opus 5.5, Opus 5, Sonnet 5, Opus 4.7/4.8low, medium, high, xhigh, max
Opus 4.6, Sonnet 4.6low, medium, high, max (no xhigh)
Opus 4.5low, medium, high
Sonnet 4.5, Haiku 4.5Not supported — omit effort

The API defaults to medium on Opus 5.5 and high on the other models that support effort. Set it explicitly when comparing models. See Anthropic's effort guide for model-specific guidance.

This can be combined with other features like structured outputs:

providers:
  - id: anthropic:messages:claude-opus-5
    config:
      effort: high
      output_format:
        type: json_schema
        schema:
          type: object
          properties:
            analysis:
              type: string
          required:
            - analysis
          additionalProperties: false

Structured Outputs​

Structured outputs constrain responses to a JSON schema and are supported by current Claude models. Promptfoo maps config.output_format to the API's output_config.format and adds the structured-outputs-2025-11-13 beta flag to the anthropic-beta header.

JSON Outputs​

Add output_format to get structured responses:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      output_format:
        type: json_schema
        schema:
          type: object
          properties:
            name:
              type: string
            email:
              type: string
          required:
            - name
            - email
          additionalProperties: false

You can also load the entire output_format from an external file:

config:
  output_format: file://./schemas/analysis-format.json

Nested file references are supported for the schema:

analysis-format.json
{
  "type": "json_schema",
  "schema": "file://./schemas/analysis-schema.json"
}

Variable rendering is supported in file paths:

config:
  output_format: file://./schemas/{{ schema_name }}.json

Strict Tool Use​

Add strict: true to tool definitions for schema-validated parameters:

providers:
  - id: anthropic:messages:claude-sonnet-5
    config:
      tools:
        - name: get_weather
          strict: true
          input_schema:
            type: object
            properties:
              location:
                type: string
            required:
              - location
            additionalProperties: false

Limitations​

Supported: object, array, string, integer, number, boolean, null, enum, required, additionalProperties: false

Not supported: recursive schemas, minimum/maximum, minLength/maxLength

Incompatible with: citations, message prefilling

See Anthropic's guide and the structured outputs example.

Model-Graded Tests​

Model-graded assertions such as factuality or llm-rubric will automatically use Anthropic as the grading provider if ANTHROPIC_API_KEY is set and OPENAI_API_KEY is not set.

If both API keys are present, OpenAI will be used by default. You can explicitly override the grading provider in your configuration.

Claude Pro/Max subscribers without a separate Anthropic Console key can wire up llm-rubric through a local Claude Code session by pointing the grader at anthropic:messages:<model> with apiKeyRequired: false:

defaultTest:
  options:
    provider:
      id: anthropic:messages:claude-sonnet-5
      config:
        apiKeyRequired: false

See Authenticating via a Claude Code session above for how the credential is loaded and what beta headers Promptfoo sets.

Because of how model-graded evals are implemented, the model must support chat-formatted prompts (except for embedding or classification models).

You can override the grading provider in several ways:

  1. For all test cases using defaultTest:
promptfooconfig.yaml
defaultTest:
  options:
    provider: anthropic:messages:claude-sonnet-5
  1. For individual assertions:
assert:
  - type: llm-rubric
    value: Do not mention that you are an AI or chat assistant
    provider:
      id: anthropic:messages:claude-sonnet-5
      config:
        effort: low
  1. For specific tests:
tests:
  - vars:
      question: What is the capital of France?
    options:
      provider:
        id: anthropic:messages:claude-sonnet-5
    assert:
      - type: llm-rubric
        value: Answer should mention Paris

Additional Capabilities​

  • Caching: Promptfoo caches previous LLM requests by default.
  • Token Usage Tracking: Provides detailed information on the number of tokens used in each request, aiding in usage monitoring and optimization.
  • Cost Calculation: Estimates each request from its model and reported token usage. For Claude 4.6 and later, U.S. inference adds 10% to token costs, including cache reads and writes. The response's usage.inference_geo takes precedence over config.extra_body.inference_geo. Explicit cost, inputCost, or outputCost overrides take precedence over this surcharge. See Anthropic pricing.

When using the Anthropic disk response cache across runs, assign each provider a distinct, non-secret label. Promptfoo uses that stable label to isolate cached responses without persisting API-key or OAuth-token fingerprints. Unlabeled providers receive an ephemeral per-instance namespace, which safely preserves repeated calls within a run without reusing responses across tenants or processes. Requests with custom headers bypass the disk response cache because those headers may contain tenant credentials.

providers:
  - id: anthropic:messages:claude-sonnet-5
    label: tenant-a
    config:
      apiKey: '{{env.ANTHROPIC_API_KEY_A}}'
  - id: anthropic:messages:claude-sonnet-5
    label: tenant-b
    config:
      apiKey: '{{env.ANTHROPIC_API_KEY_B}}'

See Also​

Examples​

We provide several example implementations demonstrating Claude's capabilities:

Core Features​

  • Tool Use Example - Shows how to use Claude's tool calling capabilities
  • MCP Example - Connect Claude to a Model Context Protocol server and let it execute the discovered tools
  • Structured Outputs Example - Demonstrates JSON outputs and strict tool use for guaranteed schema compliance
  • Vision Example - Demonstrates using Claude's vision capabilities

Model Comparisons & Evaluations​

Cloud Platform Integrations​

Agentic Evaluations​

  • Claude Agent SDK - For agentic evals with file access, tool use, and MCP servers

For more examples and general usage patterns, visit our examples directory on GitHub.

Last updated on by mldangelo-oai