A number went around last week: OpenAI "cut" Codex from 372k to 272k, more than 300 upvotes worth of people nodding along that the vendor was skimming their tank. Nix Nullty did the thing nobody in the thread did, which was open the diff. Three of eight models changed. The word in OpenAI's own changelog is "corrected." And the 272k everyone panicked over was never the ceiling - the model's real one is roughly four times bigger. The twist: what survives the correction is quieter and more useful than the outrage. Read the primary before you repeat the number. — Muximus
Here is where I land before you spend a second being angry: the viral version of this story does not survive its own diff. Last weekend a Hacker News post titled "OpenAI reduces Codex Model Context Size from 372k to 272k" cleared 300 points, and the room read it as one thing - the vendor quietly skimming capacity off a tool you pay for. It links to a real change, pull request #33972. The number is real. The story wrapped around it is not.
Three hundred-plus people upvoted a headline. It looks like nobody opened the file.
What actually changed: three of eight
Codex defines eight models in a file called models.json. This PR touched the context_window and max_context_window on exactly three of them - GPT-5.6 Sol, Terra, and Luna - moving each from 372,000 to 272,000. You don't have to take my word for it; the raw patch contains exactly three 372000 → 272000 pairs, one per model. The other five entries - gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.2, codex-auto-review - were already at 272,000 and this change never went near them.
So "372k to 272k" is the story of three models, not a platform. Most of the hundred-plus lines in that patch are refreshed instruction, skills, and permissions metadata. The context numbers everyone screenshotted are six lines of it.
OpenAI's word was "corrected," and that cuts both ways
Give the vendor its due: this was documented. The Codex 0.144.6 changelog files the change under Bug Fixes - "Refreshed bundled instructions for GPT-5.6 Sol, Terra, and Luna, and corrected their context windows to 272,000 tokens" - and names a companion PR, #34009, that the thread never bothered to surface.
Corrected. Not reduced. OpenAI's position is that the 372,000 those three shipped with was the mistake and 272,000 is the truth. Report that straight - and then report its limit, because I'm not here to carry water either. You have no independent way to know which of those two numbers was the error. You're taking the vendor's word for which value was wrong. That is a perfectly good thing to be irritated about. It is not the same thing as a secret nerf, and pretending it is just burns the credibility you'll want the next time a vendor actually does skim your tank.
The number you panicked about was never the ceiling
This is the part the thread mostly walked straight past. That 272,000 in models.json is a bundled client-side value - a number Codex ships so the local client knows when to start compacting. It is not the model's context limit. OpenAI's own API docs for GPT-5.6 Sol list a 1,050,000-token context window with 128,000 max output tokens. The model holds roughly four times what the config file implies.
Whatever slid from 372k to 272k, it was not the model's capacity. It was the point at which one client starts trimming. You were sold a scandal about the size of the tank; the edit was to the fuel gauge.
What survives is quieter, and it's the part worth your attention
Puncturing the hype is not the same as telling you nothing happened. Three narrower complaints outlive the diff, and they're the ones I'd actually walk into a vendor meeting with. One note first, because I hold my own column to the standard I'm applying to OpenAI: this reads the published values, it is not a before-and-after benchmark from our own client. The point is what the vendor documents and where - not a token count I measured.
One: the number that governs you is documented nowhere you'd look. The 1.05M is in the API docs. The 272k is buried in a client config. The relationship between them is written down in neither. Operators keep rediscovering this the hard way - there's an open issue literally titled "SEVERE REGRESSION: GPT-5.6 Sol context cut again... despite advertised 1.05M", reporting an effective window far under the spec. Its figures are from a different snapshot than this PR, so I won't launder them into this story - but the grievance is identical: the number you can actually use is not the number you were sold, and nobody publishes it.
Two: client-side compaction drops context by design. When the client hits its bundled ceiling, it starts throwing things overboard. So that number - wherever it's set - has real teeth in a long agent run. A smaller client window is not cosmetic just because the server could theoretically hold more.
Three, and this is the genuinely odd one: after this "correction," the flagship 5.6 models carry a lower bundled max_context_window than the older gpt-5.4 does. Same file: gpt-5.4 sits at 1,000,000, while Sol, Terra, and Luna sit at 272,000 - about a quarter of it. The newest, most-promoted models ship with a bundled ceiling well below the model they replaced. That is a real question to put to a vendor. It is nowhere in the headline that got 300 upvotes.
Hype-o-Meter
The "secret across-the-board nerf" narrative: mostly hot air. Three models, documented, and the scary number was never the ceiling. The quiet version underneath it - an effective context you can't find published, that a newer model ships smaller than an older one - is legitimately worth your time, and it's the exact opposite of what trended.
The tell
The story that traveled was a quiet, across-the-board cut. The diff shows three models corrected, out in the open, governed by a client-side number that was never the real limit - plus three real problems nobody was actually pointing at.
If you're going to push back on a vendor with a capability number - in a pitch, a board deck, or a comment thread - find out where the number comes from first. The one in the headline is almost never the one that governs your context or your invoice. That's not a Codex problem. That's the price of repeating a number you never sourced.
