Extended Thinking
Claude API · Anthropic
Foundation Model & API — Anthropic (Claude)
Anthropic evals — Extended Thinking (relift v3 InfraRed)
About Anthropic
Anthropic is an AI safety company and the maker of Claude. Its API exposes the Claude model family (Opus, Sonnet, Haiku) with tool use, prompt caching, extended thinking, batch processing, vision, the Files and Memory tools, and the Claude Agent SDK.
Sample tests· showing 3 of 9
| # | Input | Expected behavior | Check |
|---|---|---|---|
| 01 | Agent enables extended thinking with thinking={type:'enabled', budget_tokens:10000}, max_tokens=16000. | Both fields are required when enabling. budget_tokens reserves thinking capacity within max_tokens (text output gets max_tokens - thinking tokens used). Verify thinking + text totals stay under max_tokens. | Pass / FailAi Platformhigh |
| 02 | Response contains redacted_thinking blocks (server-side encrypted) instead of plain thinking. | Preserve redacted_thinking blocks verbatim — including the redacted data — in subsequent assistant turns sent back to the API. Do NOT modify, decode, drop, or log their contents. | Pass / FailAi Platformcritical |
| 03 | Agent sets thinking enabled and temperature=0.2 expecting deterministic thinking. | Extended thinking requires temperature=1 per docs. Set temperature=1 (or omit) when thinking is enabled. Other temperatures cause invalid_request_error. | Pass / FailAi Platformmedium |
How this eval is graded
Grade against expected.ideal_behavior and expected.rubric. Per-criterion pass requires mean >= 4.0 and no criterion below 3.
Rubric criteria
- Anthropic
- Ai Platform
- Extended Thinking
Recommended for
Works with
Related evals
Run this eval in your workspace
Connect your data, configure thresholds, and review results with your team.