1. A lap is one call. Its prompt is the whole context so far plus the new input of the lap (tool results, user turn).
2. The reply is billed at the output price and appended to the context. Its size is the base output times the effort multiplier.
3. Cached input on: the prompt already sent on the previous call is read at the cache-read price, and everything new since then (last reply and new input) is written at the 5-minute cache-write price. Cached input off: the whole prompt is billed at the base input price.
4. Stints are deduced, not chosen. Before a lap, if the prompt would cross the pit threshold, the agent pits. A pit costs one extra call that reads the context and writes the handoff as output. The next stint starts from the handoff seed with a cold cache.
5. Haiku 5.5 is tiered by prompt size: each call uses the rates for prompts up to 100k tokens or over 100k tokens.
6. Not modelled: cache expiry between laps (laps are assumed to come within 5 minutes), system prompt and tool definitions, batch or regional pricing, and answer quality as the context grows.

Each model's default level is rescaled to 1.00, so Sonnet 5.5 and Fable 5.1 (default high) divide these values by 1.38. Cross-check against the other published figures: Sonnet 5.5 medium costs about 59% of high (this table gives 72%), and Fable 5.1 low costs $4.48 against $6.89 for high (65%; this table gives 38%). Source: the SWE-bench Pro subset in the effort docs. Prices: prices as of 2026-10-28.