OpenAI rolls Codex context back to 272k days after pushing it to 372k — three config changes, three rollbacks in one week
TL;DR
OpenAI rolls Codex context back from 372k to 272k just two days after raising it — GPT-5.6 Sol picks up ~10% inference savings.
OpenAI product lead Tibo Sottiaux announced early July 13 that Codex and ChatGPT Work's context ceiling is rolling back from 372k to 272k, just two days after the 272k → 372k raise caused billing to overshoot user expectations. The 372k tier will return in the coming days, but the public rollback itself is unusual.
The batch of updates covered three more items. GPT-5.6 Sol finished an inference optimization pass — the same task now consumes fewer tokens, which effectively hands users about 10% more usage. The reasoning effort (internal name "juice") experiment is reverted after high / xhigh effort caused auto-review to spawn too many agents and hurt throughput. And the 5-hour usage cap remains paused for Plus, Business, and Pro tiers, with reset cadence shifted to hourly batches.
Backdrop: OpenAI shipped aggressive config changes on the Codex side all week — 372k context on Monday, juice tuning on Wednesday, auto-review changes on Thursday — every one of which now sits in a rolled-back state. Tibo's announcement effectively concedes the weekly-iteration cadence billed users.
Active users on the ChatGPT platform crossed 600 million last week. During the 5-hour cap pause, quota resets have moved to hourly batches — users who hit the cap in the last hour cycle through fresh quota on a rolling basis.
Win the bet, and OpenAI's 372k context + tuned juice becomes the retention lever on paying users. Lose it, and this week's three public rollbacks are the trust hit that follows.
via Tibo Sottiaux on X
The batch of updates covered three more items. GPT-5.6 Sol finished an inference optimization pass — the same task now consumes fewer tokens, which effectively hands users about 10% more usage. The reasoning effort (internal name "juice") experiment is reverted after high / xhigh effort caused auto-review to spawn too many agents and hurt throughput. And the 5-hour usage cap remains paused for Plus, Business, and Pro tiers, with reset cadence shifted to hourly batches.
Backdrop: OpenAI shipped aggressive config changes on the Codex side all week — 372k context on Monday, juice tuning on Wednesday, auto-review changes on Thursday — every one of which now sits in a rolled-back state. Tibo's announcement effectively concedes the weekly-iteration cadence billed users.
Active users on the ChatGPT platform crossed 600 million last week. During the 5-hour cap pause, quota resets have moved to hourly batches — users who hit the cap in the last hour cycle through fresh quota on a rolling basis.
Win the bet, and OpenAI's 372k context + tuned juice becomes the retention lever on paying users. Lose it, and this week's three public rollbacks are the trust hit that follows.
via Tibo Sottiaux on X
