The 258K ceiling
OpenAI sells GPT-5.5 as its best model for working with large amounts of text at once, yet inside its own Codex coding agent the same model quietly runs at about a quarter of its advertised capacity.
OpenAI's GPT-5.5 can hold a million tokens (the chunks of text a model reads and writes) in view at once through its developer interface. That capacity is the model's working memory for a task. Load the identical model into Codex, OpenAI's flagship coding agent, and live sessions report an effective working memory of 258,400 tokens against a published figure of 400,000. Developers worked the number out from Codex's own settings: it sets GPT-5.5 to 272,000 tokens and shaves off 5 percent. 272,000 times 0.95 is 258,400, the exact figure sessions display.
The Codex model catalog caps GPT-5.5 at 272,000 tokens, so local overrides are clamped or ignored. — openai/codex issue #19464
The sharper detail is the direction of travel. The previous model, GPT-5.4, let Codex users stretch the working memory up to a million tokens; upgrading to the model OpenAI markets as stronger at handling more text shrank the usable memory inside the same tool by roughly four times. And the auto-compaction meant to rescue a long session by summarizing its earlier parts is aimed at the advertised size, not the real one, so it fires too late and often fails outright, dropping the thread mid-task.
This is a narrow grievance with a wide lesson. The people who hit it are the heavy users: multi-file refactors, long debugging loops, large repositories, the sessions a coding agent is supposed to be best at. It surfaced not through a launch or a benchmark but through a cluster of GitHub issues where developers audited the config and did the arithmetic. The gap that matters is between the number a vendor advertises for a model and the number its own product silently enforces.
The lenses
The facts
Concepts
How this connects
Tap a node to open it