Claude Task Budgets: Cap Agent Token Spend in Python

The new-model wave that landed at the end of June 2026 (Claude Sonnet 5 on the 30th, Opus 4.8 already shipping, OpenAI's GPT-5.6 preview) all point the same direction: agents that run for minutes or hours, calling tools, thinking, and re-planning. That is great for capability and terrifying for your invoice. A single agentic loop that goes sideways can quietly spend a million tokens before you notice.

Anthropic's answer is a beta feature called task budgets. Instead of hoping your agent behaves, you hand it a token budget for the whole loop. The model watches a live countdown and paces itself: it prioritizes the important work early and wraps up gracefully as the budget runs low, rather than getting chopped off mid-action by a hard limit.