OpenAI Boosts Codex Default Context to 500k Tokens
OpenAI bumped the default Codex context for GPT-6.1 Sol to 500k tokens. That means fewer forced summaries and more of your code in view.
Tools · Source: TSN Media
What happened
OpenAI is currently on day seven of a 28-day shipping streak. The latest update targets developers directly. Codex lead Thibault Sottiaux announced on X that the default context length for GPT-6.1 Sol in Codex is now 500,000 tokens. This represents an 80 percent jump from the previous 272,000 token default. The change applies when using the model through Codex and ChatGPT Work.
This update tackles a major pain point for engineers. A context window dictates how much text, code, and chat history a model can hold in view at once. When a long coding session fills this window, Codex compacts the history. It shrinks older material into a summary so the work can continue. This compaction process often loses crucial detail. Sottiaux noted that a larger default helps reduce the occurrence of these compactions.
OpenAI also rolled out composer predictions in beta for Pro users. The composer is the message box where you interact with the model. Codex can now suggest your next message based on your conversation history and how you talk to it. You simply press Tab to accept a suggestion and edit it if needed. Meanwhile, ChatGPT always-on agents, known as dots, received an upgrade. Users can now create dots directly from their mobile phones on iOS and Android. These dots can start work in Codex and follow up on existing threads.
Key facts
- 500,000 — New default token context length for GPT-6.1 Sol in Codex
- 272,000 — Previous default token context length
- 1,000,000 — Maximum context length users can manually set for GPT-6.1 Sol
- 80% — Increase in the default context length
Why it matters
A 500,000 token default changes how you build software with AI. You can now dump massive repositories, extensive documentation, and long chat histories into Codex without the model losing the plot. The model does not inherently get smarter with this update. Instead, the forgetting line moves much further away. Users can also manually adjust the context up to the model absolute limit of one million tokens. One million tokens is roughly equivalent to two thousand pages of code. Sottiaux mentioned that OpenAI plans to push the default context even higher in the future.
The second-order effect is a massive shift in developer momentum and workflow automation. OpenAI is aggressively removing friction from the coding process. Features like composer predictions anticipate your needs before you even type them. Mobile dot creation lets you start a project idea on your phone while commuting and have the agent execute it on your connected desktop IDE. The barrier between having an idea and executing it is collapsing rapidly. Builders who leverage these expanded memory limits and agent hand-offs will ship products significantly faster than those who do not.
For builders
Dump massive repositories into your prompts
You now have 500,000 tokens by default before Codex starts summarizing your chat history. You can manually push this limit to one million tokens if needed. This saves engineers time by keeping massive codebases and documentation in active memory much longer.
Speed up workflows with composer predictions
Pro users get next-message suggestions in the Codex desktop app for local tasks. Accepting a prediction does not automatically send it, giving you time to edit. This feature is free during the beta and generating predictions does not consume your Codex usage limits.
Start local coding tasks from your phone
You can now create always-on agents, called dots, from the ChatGPT mobile app on iOS and Android. These dots can hand off work directly to Codex on a connected computer. Founders can spin up project ideas on the go and have the agent execute them locally.
My take
I see this update as OpenAI simply moving the forgetting line further away. They are not teaching Codex to remember everything forever. But giving us a half-million tokens by default proves the bottleneck is no longer model capability. The real edge goes to founders who know how to feed massive context into these tools without getting overwhelmed.
Original reporting: TSN Media. This is my rewrite and opinion.