OpenAI Codex quietly slashes GPT-5.6 context from 372K to 272K | adds "don't rm -rf $HOME" to system prompt
TL;DR
OpenAI's July 19 Codex update quietly capped effective GPT-5.6 input context at 272K (down from 372K); requests over the line are billed at 2x input / 1.5x output for the whole call. The raw 1.05M model spec is unchanged. The same update also added a preset system-prompt instruction blocking rm -rf $HOME.
The Codex update OpenAI shipped on July 19 caps the effective GPT-5.6 context window at 272K input + 128K output, down from 372K + 128K. The raw model spec is unchanged at 1.05M context — what Codex is doing is aligning its bundled catalog to the price-tier line: requests crossing 272K get billed at 2x input pricing and 1.5x output pricing for the entire call, not just the overage.
The problem is that most AI coding harnesses were built around 372K, and many don't warn users before crossing the new line. Teams running long-context coding sessions are about to see next month's bill jump without explanation. OpenAI did not publish a blog post — the change lives only in Codex PR model metadata, blended into a normal release, with no deprecation notice and no migration guide.
A quieter change in the same update: the Codex team added a preset system-prompt instruction that forbids the model from running rm -rf $HOME. Using a system-prompt guardrail like this only makes sense if full-access mode has already erased real users' home directories. No overlayfs-style filesystem isolation is apparently on the roadmap.
via AI Weekly / 蓝点网
The problem is that most AI coding harnesses were built around 372K, and many don't warn users before crossing the new line. Teams running long-context coding sessions are about to see next month's bill jump without explanation. OpenAI did not publish a blog post — the change lives only in Codex PR model metadata, blended into a normal release, with no deprecation notice and no migration guide.
A quieter change in the same update: the Codex team added a preset system-prompt instruction that forbids the model from running rm -rf $HOME. Using a system-prompt guardrail like this only makes sense if full-access mode has already erased real users' home directories. No overlayfs-style filesystem isolation is apparently on the roadmap.
via AI Weekly / 蓝点网
