GPT-5.6 can use about 8× more tokens than 5.5 because its harness splits batch jobs into repeated tool calls

breath_mirror · x · 2026-07-27

The post argues that GPT-5.6 can burn roughly 8× more tokens than GPT-5.5 because the new harness often splits one batched operation into many separate tool calls and re-reads the full context each time.

The core point is that the cost jump appears to come less from the model itself and more from the agent harness design, which turns efficient batch work into repeated context-heavy loops.

Original post →

GPT-5.6 can use about 8× more tokens than 5.5 because its harness splits batch jobs into repeated tool calls

More from coding & agent