OpenAI has quietly shrunk the amount of context its Codex coding tool feeds to GPT-5.6, and developers who rely on long agentic sessions are unhappy. The default configured input context window fell from 372,000 to 272,000 tokens — a cut of about 27% — made through GitHub pull request #34009.
What changed
The context window is the working memory of a coding agent: the span of code, instructions and prior conversation the model can hold at once. Trimming it by roughly a quarter means Codex can keep less of a session in view before it must summarize and discard earlier material — a process known as compaction.
Why developers noticed
The practical effect lands hardest on long, multi-file tasks. "Less memory per session means the AI agent forgets earlier parts of a long coding session sooner," analyst Pareekh Jain said. Another, Muskan Bandta, noted that "a lot of developers are saying their sessions now spend more time on compacting than actually working" — the tool busy managing its own memory instead of writing code.
No stated rationale
OpenAI has not publicly explained the change, leaving developers to infer the motive. The obvious candidate is cost: serving a larger context window consumes more compute per request, so quietly reducing it lowers the expense of running Codex at scale — the kind of trade-off labs increasingly make as inference demand strains capacity.
Adapting around it
Analysts say the fix is workflow, not complaint. Teams should "design workflows that proactively manage context," analyst Amit Jena said — breaking work into smaller tasks, relying more on retrieval mechanisms and monitoring context consumption. It is a reminder that the usable capacity of an AI tool can shift without notice, and that building around a fixed context budget is risky.