Cut your agent bill

Eight tactics that work, tied to the actual price sheets.

1. Route by difficulty

Use a cheap model (GPT-6 Luna at $0.10/$0.50, Gemini 3.5 Flash-Lite, Claude Haiku 5.5) for search, summaries and boilerplate; save frontier models for planning and hard debugging. Output tokens cost 5x input on most models listed here.

2. Make the cache work for you

Cached input is 10x+ cheaper on most APIs (e.g. GPT-6.1 Sol $2.00 → $0.10). Keep system prompts, tool definitions and repo context stable and at the start of the prompt so they hit cache.

3. Watch long-context thresholds

Several APIs reprice the whole request above a threshold: Grok above 200K prompt tokens, OpenAI above 272K, Gemini 3.1 Pro above 200K. Compact or summarise before you cross it.

4. Use batch/flex for anything async

OpenAI Batch/Flex and Gemini Batch are about 50% off standard. Evals, codebase indexing and nightly refactors don't need real-time answers.

5. Clear context between tasks

Agent sessions resend history every turn. Starting fresh for each task (/clear in Claude Code, new chat in Cursor) is the cheapest optimisation there is.

6. Know your plan's reset clock

Claude and ChatGPT Plus use rolling 5-hour windows plus weekly caps; Cursor's pool refreshes weekly; Copilot resets monthly; Windsurf/Devin daily + weekly. Plan heavy sessions right after a reset.

7. Set hard spending caps

Cursor (3rd-party models at API price), Copilot paid usage, Claude usage credits and Devin on-demand credits all bill past the quota. Set a budget alert before you start a long run.

8. Price changes are scheduled — read them

Gemini 4 Argon doubles to $4/$20 on Dec 1, 2026; Gemini 3.8 Flash doubles on Jan 1, 2027; GPT-5.6 Sol promo pricing runs at least to Nov 21, 2026. Subscribe to get alerted.