OpenAI Codex Cuts GPT-5.6 Context Window From 372K to 272K

A one-file JSON change merged into OpenAI's Codex repo this weekend is worth reading closely, because the number it edits is a billing boundary dressed up as a spec. PR #33972, titled 'Backport refreshed bundled model metadata to 0.144,' was merged July 18 by sayan-oai. According to the summary of the diff, it refreshes bundled instructions for GPT-5.6 Sol, Terra, and Luna and corrects their reported context windows to 272,000 tokens. The prior metadata reported a 372,000-token window with a 95% effective-context multiplier, which yielded an effective window of about 353,400.

The reason 272K matters is what OpenAI's own model documentation says for this family: 'prompts with >272K input tokens are priced at 2x input and 1.5x output for the full session for standard, batch, and flex.' The raw GPT-5.6 spec is unchanged at a 1,050,000 context window with 128,000 max output tokens, so what Codex is really doing here is aligning its bundled catalog to the price-tier line, so that default sessions do not silently cross into premium pricing on long requests.

A prior thread on the same repo, issue #32486, had already flagged the exposure: with the 372K figure exposed to clients, roughly 81,400 tokens sat inside the higher-cost band before any warning fired, and the default auto-compaction did not stop sessions from tripping it. The corrected metadata narrows that gap by design, and the same issue notes that a manual `model_auto_compact_token_limit = 270000` was the workaround until this landed.

The honest caveat is what the PR notes do not spell out in the retrieved material. There is no confirmation about whether users on the stable 0.144.x line will see an in-client warning when this lands, whether the 2x/1.5x multiplier is scoped only to fresh input or also to cached tokens, or how third-party Codex harnesses that hard-coded the older 372K figure will be nudged to update.

For teams shipping Codex-based agents, the practical read is to audit auto-compaction defaults this week and stop budgeting against the older number. The quieter competitive read is that whichever framework or vendor surfaces the 272K tripwire first in its own UI gets to sell 'no surprise bills' against a pricing tier OpenAI itself designed.