OpenAI has opened up a 1 million-token context window in Codex for GPT-5.6 Sol to ChatGPT subscribers, a change that lets the AI coding agent hold far more of a codebase in memory during a single session.
Thibault Sottiaux, who leads Codex engineering at OpenAI and posts online as Tibo, announced the update on X on August 16. The setting was previously limited to developers using API keys. It now also works for users on ChatGPT Plus, Pro, Business, and Enterprise plans.
What Changed
Sottiaux published the exact configuration needed to turn on the larger window. Users can edit the ~/.codex/config.toml file and set the model to GPT-5.6 Sol, then manually raise the context window to 1,000,000 tokens with an automatic history-compaction limit of 900,000 tokens.
GPT-5.6 Sol carries a documented context window of roughly 1.05 million tokens. Until this week, Codex users signed in through ChatGPT could not access that full capacity. Only API customers could push past the smaller default window built into the Codex app and CLI.
A bigger context window means Codex can keep more code, command output, and conversation history active before it starts summarizing or discarding older material. That matters most for large-scale refactoring, multi-file debugging, and long agentic sessions that would otherwise force the tool to compress earlier steps.
Why OpenAI Made the Change
The update follows weeks of developer frustration over Codex’s context limits. In mid-July, users on OpenAI’s GitHub repository for Codex reported that the tool’s effective context window had shrunk sharply, from roughly 353,000 tokens down to about 258,000 tokens, even though GPT-5.6 Sol’s published specification promised over 1 million tokens.
Multiple GitHub issues asked OpenAI to restore or make optional a larger context profile. One developer said the earlier configuration let the model hold a bigger working set of code before it had to compress older context.
Analysts said the tighter window mainly affected heavier workloads. Pareekh Jain, principal analyst at Pareekh Consulting, told InfoWorld the change could impact large codebases, repository-wide refactoring, and long-running sessions, even if routine fixes were largely unaffected.
This week’s update gives developers a way to opt back into the model’s full advertised capacity, though it does not change what Codex uses by default.
OpenAI’s Caution on the Setting
Sottiaux was clear that the expanded window is not meant to become the standard setting for most users. In his post, he noted that OpenAI had tuned the default context limit for a balance of speed and cost, adding: “Have fun, but also know that we tuned the default carefully!”
Other OpenAI staff who replied to the post echoed that caution, suggesting most coding tasks do not need the full 1 million-token budget and that enabling it by default could hurt response speed on simpler jobs.
Developers testing the setting also flagged a practical cost. Sessions that use the full window can consume a much larger share of a user’s weekly usage allowance, since GPT-5.6 Sol is billed per token and larger prompts cost more to process.
Who This Affects
- Developers working on large, established codebases that need broader context during refactors
- Teams running long agentic coding sessions across many files
- Codex users who previously needed an API key to access the model’s full context capacity
- Anyone on a metered ChatGPT or API plan who should watch usage before turning the setting on for routine work
Background on GPT-5.6 Sol
OpenAI introduced the GPT-5.6 family, made up of Sol, Terra, and Luna, on July 9. The naming system replaced OpenAI’s earlier version-number scheme, with Sol positioned as the flagship model for demanding coding and agentic work, Terra as a mid-tier option, and Luna built for faster, lower-cost tasks.
Sol is priced at $5 per million input tokens and $30 per million output tokens through the API. OpenAI has described it as its strongest coding model to date, and the company merged its Codex and ChatGPT desktop apps around the same release.
Codex’s user base grew quickly after the July launch, and OpenAI has made several follow-up adjustments to usage limits and backend performance since then, in response to developers reporting that agentic sessions were consuming allowances faster than expected.
Pakistan Context
Pakistan’s IT and freelance software sector has increasingly relied on AI coding assistants like Codex, GitHub Copilot, and Claude Code as part of daily development work. The Pakistan Software Export Board reported IT and IT-enabled services exports of $4.6 billion for FY2025-26, with freelancers contributing more than $1.1 billion of that total, a 20% year-on-year increase.
For Pakistani developers working on outsourced or freelance projects involving large codebases, a bigger context window in tools like Codex can reduce the need to break large repositories into smaller chunks during AI-assisted work, though the usage-cost trade-off Sottiaux flagged applies equally to users on metered plans in Pakistan.
What Happens Next
OpenAI has not said whether the 1 million-token window will become a standard, opt-in toggle inside the Codex interface rather than a manual configuration file edit. The company has also not confirmed whether Terra or Luna will receive similar expanded-context options. OpenAI typically rolls out such settings gradually and has made repeated context and usage-limit adjustments to Codex since GPT-5.6 Sol’s July launch.
Leave a Reply