New tool versions for code execution, web search, and web fetch

What's new

• code_execution_20260521: surfaces the 90-second per-cell limit in the tool description so Claude can budget long-running cells
• web_search_20260318: adds a response_inclusion parameter so the agent can explicitly drop consumed result blocks
• web_fetch_20260318: same response_inclusion mechanism, for leaner multi-step agentic workflows
• No beta header required: all three are GA, not experimental

Why it matters

For developers building agents on the Claude API, response_inclusion is a real optimization — long conversations with many web_search / web_fetch calls fill the context window fast. Trimming consumed blocks without losing citation capability saves both tokens and latency. The per-cell time hint on code_execution helps Claude decide whether to split or parallelize a long task.

How to try it

Set the tool version explicitly in your API request (e.g. code_execution_20260521). No beta header needed. Older tool versions stay functional for compatibility.