What does one coding-agent session actually cost?

Cache math: 70% of input tokens billed at the cache-read price (≈10% of input) where the provider supports prompt caching; models without cache pricing use full input price. Estimates use list prices verified 2026-10-21.

Why input tokens dominate agent bills

A chat request sends your question once. An agent request re-sends everything: system prompt, tool definitions, file contents, and the whole conversation — every single call. 100 calls × 10K input tokens = 1M input tokens, which at frontier pricing ($2–10/1M) is 10–50× the output cost. Three levers cut it:

• Prompt caching — stable prefixes (system prompt, tool defs) hit cache at ~10% of input price. See the slider above.
• Cheaper input pricing — models like DeepSeek or Gemini Flash price input aggressively; rank by your exact mix above.
• Batch processing — non-interactive steps (test runs, doc summarization) can use batch APIs at ~50% off. See batch pricing reality check.