OpenAI Codex capped at 272k–400k

OpenAI Codex is running into a very specific developer complaint — not that it can’t code, but that the packaging around long-context coding feels muddy. The argument flared because OpenAI’s current Codex pricing pages emphasize rate limits and credits, while the underlying model docs quietly show where long-context costs kick in. That matters if you’re using coding agents on big repositories, where context window size stops being a benchmark vanity metric and starts being a workflow limit. The story this week is basically a mismatch between how Codex is sold and how power users think about value. (developers.openai.com) ### What are developers actually upset about? They’re upset that paying more for Codex does not obviously buy a bigger context window. OpenAI’s Codex pricing says Plus, Pro, Business, and Enterprise plans include Codex, and Pro mainly gets you more usage — 10x or 20x Plus depending on tier and promos. That is useful, but it is not the same thing as “this model can hold a much larger codebase in working memory.” For people doing agentic coding, those are different products. (developers.openai.com) ### Where does the 272K number come from? It comes from OpenAI’s GPT-5.4 model page. That page says GPT-5.4 and GPT-5.4 pro have a 1.05M-token context window, but prompts above 272K input tokens get charged at 2x input and 1.5x output for the full session. So 272K is not a hard technical ceiling in the API docs. It is the point where long context becomes meaningfully more expensive. If users are seeing Codex sessions top out a(developers.openai.com)t-layer limit on top of the raw model capability — or at least a cost wall that feels like one. That last part is an inference, but it fits the docs better than “the model itself only supports 272K.” (developers.openai.com) ### Is Codex the same thing as the API? No — and that is the source of a lot of confusion. OpenAI’s API docs describe model capabilities and token pricing. Codex is the product layer — the coding agent in ChatGPT, the CLI, the IDE extension, and related workflows. OpenAI’s help pages say Codex usage depends on plan, task complexity, codebase size, long-running tasks, and how much context the agent has to hold. So a m(developers.openai.com)s, trims, or prices that window in ways users experience as a cap. (help.openai.com) ### Why does long context matter so much in coding? Because coding agents are not just answering one prompt. They read files, keep track of earlier edits, run tests, inspect logs, and bounce across modules. A million-token window is like giving the agent a much bigger workbench — more of the repo, more prior attempts, more test output, fewer resets. Once the window gets tight, the agent has to sum(help.openai.com) reliable. (help.openai.com) ### Why are people dragging Anthropic into it? Because Anthropic is marketing the contrast very clearly. Anthropic says Claude Opus 4.6 has a(help.openai.com)pers love simple numbers. (anthropic.com) ### Is OpenAI being misleading? That’s too strong. OpenAI is pretty explicit that Pro buys more Codex usage, not unlimited everything, and that larger codebases consume more allowance. But the catch is that many developers hear “best plan for coding” and assume they are also buying the fullest version of long-context capability. OpenAI’s docs do not present that distinction very prominently. (developers.openai.com)is is less about one angry thread and more about how coding-agent products are now being judged. Raw benchmark wins still matter. But pricing clarity, usable context, and what your subscription actually unlocks matter just as much. If OpenAI wants Codex to feel premium, it probably needs to explain the gap between model context, product limits, and paid-plan entitlements a lot more plainly. (developers.openai.com)