OpenAI’s Tibo (Thibault Sottiaux) released an adjustment on October 12 that should bring relief to heavy code users: in the Codex and Work environments, the default context window for GPT-6.1Sol is suddenly increased from 272,000 tokens to 500,000 tokens, which is equivalent to expanding the model’s “short-term memory” by about 80%.
Tibo explained that although OpenAI will continue to raise this upper limit in the future, this expansion is already enough to significantly reduce the frequency of context compression being triggered—context compression refers to the awkward moment when the model’s memory cannot hold more content, and it is forced to quickly summarize and discard details. After the window becomes wider, tasks that require holding a large amount of context for a long time and are filled with complex details can be handled more easily by the model, without frequent cleaning up.
He especially emphasized that the default value for these 500,000 tokens has been thoroughly tested, and for the majority of users, it strikes the optimal balance between experience and performance—the so-called “sweet spot”. Of course, if anyone feels it isn’t sufficient or thinks it’s a waste, the manual knob can still be turned on. The maximum context length supported by GPT-6.1Sol is 1 million tokens; users can adjust it either to its maximum or back down to a lower level.
The most direct blow fell on the bill. With this window expansion, when OpenAI deducts the usage from the subscription, it will no longer charge a long context fee separately for GPT-6.1Sol—the extra money that was paid in the past due to prolonged context now gets erased directly. When a company extends the model’s memory line to 500,000 and conveniently eliminates the premium for long texts, what it wants to retain is clearly not just the developer’s code, but also the bill that would otherwise expand along with the context in the developer’s wallet.