Anthropic API: new code execution tool version exposes 90-second limit, web tools gain response_inclusion parameter

Anthropic released on June 11, 2026, a small but technically relevant set of upgrades to three API tools widely used in agentic pipelines. The changes address two problems that appear in production environments: unexpected timeout overruns during code execution, and excessive token costs in long multi-step sessions.

Code execution tool: Claude now knows about the 90-second limit

The new version of the code execution tool is labeled code_execution_20260521. The key change is deceptively simple: the tool description within the API interface now explicitly states that each cell has a 90-second execution limit.

Why does this matter in practice?

In previous versions, the limit existed but Claude had no formal information about it within the tool’s context. The consequence: when an agent triggered a long-running computation — such as iterating over a large dataset, running a complex numerical simulation, or evaluating an ML model — the cell would simply time out and the agent would receive an error with no clear explanation.

With the new version, Claude has information about the constraint from the moment it receives the tool description. It can use this during planning: split the task into smaller segments that fit within the time window, assess whether the task should be approached differently altogether, or proactively warn the user that the computation takes too long for a single-pass approach.

No beta header is needed. The change is available immediately in all production environments using the new tool version.

What is the new response_inclusion parameter and why does it save tokens?

web_search_20260318 and web_fetch_20260318 — the new versions of the web search and URL content retrieval tools — introduce a shared response_inclusion parameter.

Mechanism

In a typical agentic pipeline, an agent calls web_search, receives results, analyzes them, and moves on. In the default mode, consumed result blocks still remain in the API response and are counted as tokens returned to the client. In long sessions with many web calls, this accumulates into a significant cost.

The response_inclusion parameter allows the developer to exclude those consumed blocks from the API response — Claude has already read and used them; there is no reason to return them and charge them as output tokens.

Who does this matter for?

The change is especially useful for:

• Research agents that perform dozens of web searches and fetches in a single session
• RAG (Retrieval-Augmented Generation) pipelines that combine web fetch with an internal knowledge base
• Monitoring agents that periodically check web sources and build a synthesis

The actual savings depend on result size and number of steps, but in long agentic sessions with many web calls the difference can be substantial.

No beta header is needed here either. Both tools are available immediately for production use.

Context: small changes, concrete impact

These three updates do not introduce new capabilities — Claude could already search the web and execute code. These are engineering fixes that address concrete pain points developers reported in production: unexpected errors from timeout overruns and unjustifiably high token costs in agentic loops.

The combination of these two changes — better self-awareness about execution limits and reduced overhead on web calls — is directly aimed at developers building long-running, autonomous agentic environments, where every unnecessary token and every preventable error carries a real cost.