Restoring the Context They Cut, Nearly Doubled──OpenAI Expands Codex's Default Context from 272K to 500K and Drops Long-Context Surcharges for Subscribers
OpenAI expands Codex's default context from 272K to 500K—up to 1M, with no long-context surcharge on subscription usage. It effectively rolls back a number it had quietly cut itself weeks earlier, though concerns about quota consumption and retrieval accuracy remain.
For a coding agent, the context window (how much information it can load at once) is its working memory. When it's too small, long sessions or large repositories quickly trigger the "forgetting older material" process, forcing developers to re-explain the premises. OpenAI cut this number just a few weeks ago, and has now nearly doubled it.
Default goes from 272K to 500K, up to 1 million tokens manually
On October 12, 2026, OpenAI raised the default context value for GPT-6.1 Sol, the default model in its terminal-based coding agent "Codex" and its business-oriented "Work." The changes are as follows.
- Default raised from 272K to 500K tokens. Out of the box, it can now retain more code and conversation history.
- Can be manually raised to a maximum of 1M (1 million tokens). This leaves headroom for huge repositories and long-running autonomous work.
- No long-context surcharge on usage drawn from a subscription. In other words, using a wide context within a flat-rate plan incurs no overage charge.
This change falls on day 7 of OpenAI's "28 days of shipping sprint" (a public commitment, started on October 5, to ship one clear improvement every day or else declare a full reset). On the sprint's first day (October 6), OpenAI said it raised the default speed of GPT-6 Astra and GPT-6.1 Sol by about 50%, from 30 tokens per second to 50.
Rolling back, on its own, a number it had cut weeks earlier
This expansion takes on a different meaning in light of what came just before it. Not long ago, OpenAI had cut the default context of GPT-5.6, the previous default model on Codex, from 372K to 272K—roughly 100,000 tokens. The change went in via a GitHub pull request (#34009), and no reason was made public.
Developers voiced a string of complaints. A narrow default means compaction of the conversation starts sooner, earlier decisions drop out during long work, and the effort of re-pasting the premises grows. Because context is the agent's working memory, a tight limit creates a state where "you spend more time managing context than doing the actual work." This 272K→500K move effectively recovers that reduction and adds more on top.
Impact on workflows and your wallet
There are two places this matters in practice. First, long-running, large-scale work is less likely to break down. For refactors that span a huge codebase, or autonomous tasks that keep running even after you close your laptop, the more context that can be retained, the less rework there is.
Second, cost predictability. Normally there are situations where extremely long inputs carry an extra charge (long-context billing), but this time OpenAI says it won't apply that to subscription usage. The ability to build agents around a wide context within a flat rate is hard to ignore for teams considering unattended operation or always-on tasks. Surrounding tools have followed quickly, too: OMP, one agent platform, raised its default context for Codex models from 272K to 500K in its October 12 release (v18.10.0).
Wider isn't always better
At the same time, concerns about widening the default have surfaced. The community has pointed out that a larger default can burn through the flat-rate quota faster, and that as context gets longer the accuracy of retrieving the needed information can actually drop (the so-called "lost in the middle" problem). Even though you can narrow it manually, the wide initial value means you should watch for unintentionally consuming more than expected.
The fact that the number swung up and down over a short period also underscores a reality: the context limit is a variable that can change at the provider's discretion. The more your operation depends on long sessions, the safer it is to set explicit limits and compaction policies on your own side so that a change to the default doesn't throw your work off balance.
References: 28 days of Shipping at OpenAI (OpenAI Developer Community) / OpenAI's Codex context reduction for GPT-5.6 sparks dissatisfaction (InfoWorld) / OpenAI quietly cut Codex context window (Eden AI) / AI Tools Changelog Hub (Gradually)
