OpenAI engineer Thibault Sottiaux (Tibo) announced on the same day that the default context length of GPT-6.1 Sol in the Codex and Work environments has been increased from 272k tokens to 500k, an increase of approximately 80%.
This change means that the amount of content a model can handle in a single conversation is almost doubled—whether it comes to processing an entire code library or reading long reports, or handling complex tasks with multiple rounds and detailed elements, it becomes harder for automatic compression to occur, and conversations are less likely to be cut off midway or forced into summaries. For example, tracking the sources of terms in a contract of hundreds of pages, or having AI continuously follow up on a development project that requires continuous updates, longer windows mean fewer “forgetfulness” and less need for repeated explanations. For users with a high dependence on long-term memory, the experience improvement is quite direct, which also aligns with the industry competition regarding the "window size" of large models.
Manually it can reach 1M, and this time there is no charge for long text.
Tibo explained that 500k is the "sweet spot" for experience and performance that the team found after repeated testing for most users, but it is far from the end point: users can still adjust manually, and the maximum limit supported by GPT-6.1 Sol has reached 1M.
More importantly, there is a change on the billing side—with this adjustment, OpenAI will no longer charge for long context usage when deducting usage based on subscriptions. In the past, such long windows were often considered as expensive add-ons, but now, with the default increase and no additional fees, it indirectly lowers the threshold for heavy usage. Given that long context has become a key selling point for all companies, giving away the benefits of expansion directly to users can be seen as a practical competitive strategy by OpenAI.