Codex adds 400K context for coding
- OpenAI rolled out GPT-5.5 in Codex on April 23, positioning it as the default model for implementation, refactors, debugging, testing, and validation. - OpenAI says GPT-5.5 scores 82.7% on Terminal-Bench 2.0 and 73.1% on its internal Expert-SWE test while matching GPT-5.4 latency. - OpenAI’s docs now frame Codex around longer-running coding sessions, while API pages list 400K to 1.05M context tiers. (openai.com)
A context window is the amount of code, logs, docs, and chat a model can keep in view at once. OpenAI’s latest Codex update pushes that limit higher for coding work. (openai.com 1) (openai.com 2) On April 23, OpenAI added GPT-5.5 to Codex and said it is the recommended choice for most Codex tasks. The company named implementation, refactors, debugging, testing, validation, and knowledge-work artifacts as the main use cases. (openai.com 1) (openai.com 2) OpenAI said GPT-5.5 matches GPT-5.4 on per-token latency in real-world serving while using fewer tokens to finish the same Codex tasks. In the same release, OpenAI reported 82.7% on Terminal-Bench 2.0 and 73.1% on its internal Expert-SWE evaluation. (openai.com) For developers, a larger context window means fewer handoffs. Instead of pasting one file at a time, a model can read more of a repository, more test output, and more documentation in one pass. (openai.com 1) (openai.com 2) OpenAI’s public materials now show several context tiers rather than one single Codex number. The GPT-5.2-Codex model page lists a 400,000-token context window with 128,000 max output tokens, while the API platform page lists 1.05 million context for GPT-5.5 and GPT-5.4, and 400,000 for GPT-5.4 mini. (openai.com) (openai.com) That helps explain why some users have seen conflicting figures like 272,000 and 400,000 in different Codex surfaces. A GitHub issue in the openai/codex repository flagged that mismatch in October 2025 after the CLI showed 272K while platform docs showed 400K. (github.com) The practical effect is on bigger edits, not just longer chats. OpenAI’s Codex changelog ties recent model updates to “long-horizon” work, context compaction, and stronger performance on large code changes such as refactors and migrations. (openai.com) OpenAI is also pricing the coding stack to separate premium and cheaper tiers. The API page lists GPT-5.5 at $5 per million input tokens and $30 per million output tokens, while GPT-5.4 mini is listed at $0.75 input and $4.50 output with a 400K window. (openai.com) So the story is less a single “400K Codex” launch than a broader shift in how OpenAI is packaging coding models. Codex is being tuned for longer, more autonomous software sessions, with different model tiers carrying different context limits and costs. (openai.com) (openai.com)
