Day 12: Tibo Sottiaux adds cache reuse to Codex

The Day 12 update sits inside OpenAI's "28 days, daily improvement or full reset" program. Previous days addressed different bottlenecks: Day 7 expanded the standard context window from 272K to 500K, and Day 11 reduced capacity-shortage errors and added automatic continuation. Day 12 targets a third problem — the repeated processing of identical context — by reusing context across subagents, holding caches longer, and routing requests to machines that already hold the relevant cache.

The practical effect depends on workload shape. Tasks that share long specifications or repository rules across several subagents stand to gain the most; tasks where each subagent works on entirely separate context may see little change. Because OpenAI did not publish hit-rate or savings figures, users comparing before and after should control for model, context size, subagent count, and task difficulty rather than relying on a remaining-usage percentage alone.

Separately, the GPT-6 Astra Pro update for the Chat tab on Pro plans adds parallel agents with progress tracking and artifact review. That is a different surface from the Codex and Work cache improvements, and the two should not be read as one combined quota change.