Estimate cost of an agent run

Short answer. Cost = Σ over model calls of (input tokens × input price + output tokens × output price), plus tool costs. Agents re-send the growing context every turn, so total input tokens grow roughly quadratically with the number of turns; the context size, not the number of tasks, is what drives cost.

Estimation formula

For n turns with an average context of c tokens and o output tokens per turn:

input_tokens  ≈ n × c          (each turn re-reads the context)
output_tokens ≈ n × o
cost ≈ input_tokens × p_in + output_tokens × p_out

A 40-turn run with a 30K-token context and 500-token replies sends about 1.2M input tokens and 20K output tokens.

Ways to cut it

• Prompt caching for the stable prefix (system prompt, references).
• Trim tool results before they enter the context (ids and counts, not full bodies).
• Summarize and restart the context at milestones.
• Route simple steps to a small model.
• Cap turns and set a budget; log tokens per turn so overruns are visible.

Sources

• Provider token-usage fields in API responses (usage.input_tokens, usage.output_tokens); Anthropic prompt caching (checked 2026-09-10).