The Window That Shrank

Somewhere in the last stretch of marketing, "context window" became a headline number — the digital equivalent of megapixels or horsepower, a single figure you could put on a slide and let people gasp at. Bigger window, bigger brain, more of your codebase the model can hold in its head at once. The arrows on those slides only ever point up. That's the surface story, and everyone's fluent in it by now.

Then someone read the diff. In a routine pull request to OpenAI's Codex repository, the context limit quietly dropped from 372,000 tokens to 272,000. A 27% haircut. No blog post, no changelog fanfare, no "we've made some changes to better serve you." Just a number, edited in a file, shipped under the fold where the marketing team doesn't look.

I want to be fair, because the boring explanation is probably the true one. Context windows are expensive. Every token you let the model consider is memory reserved, compute burned, latency added, margin shaved. A hundred thousand tokens of headroom multiplied across every request is a real infrastructure bill, and someone on the ops side almost certainly looked at that bill and made a sensible call. This is not a scandal. Nobody lied.

And be more careful still, because I can't actually tell you what that number was. Maybe 372k was the real ceiling and someone shaved it. Maybe 272k was always the true limit and 372k an aspirational value that had been quietly throwing errors — a drop that's a fix, not a cut. Maybe the split changed, more room reserved for the model's output and less for your input, the total untouched. Maybe 372k was never a figure anyone put in front of you at all — just a default in a config file, load-bearing for nothing you were promised. I don't know. And that's not a hedge. That's the story.

Because here's what you can't do: you can't measure your own context window. There's no gauge. If it shrank, you'd experience the shrinkage as the model "getting dumber today" — and you'd blame the model, or blame yourself, never the config line you were never shown. The number they demo is a capability: the most the system could do, on the good day, for the slide. The number you're handed is a provisioning decision: what it'll actually be given on a median Tuesday, when the constraint is a budget and not a benchmark. Those are different kinds of number, and only one of them ever makes the keynote.

The uncomfortable part is that this is the honest version of the industry. Somewhere a number went up in a marketing deck the same quarter it went down in a config file, and both are true, and no one has to reconcile them because they're told to different audiences. You'll hear about the expansion. The contraction lives in a pull request that a stranger happened to read.

And that last part is the whole reason we can talk about this at all: a stranger could read it. Codex is open source. There was a diff to open, a file to catch moving. That's not the rule — it's the exception that measures how blind we usually are. The overwhelming majority of these adjustments happen in closed weights and closed servers, where there is no diff, no file, no number to watch change. This one was catchable. Almost none of them are.

Watch the diffs, not the demos. The demos are written for you; the diffs are just written. And where there's no diff — which is nearly everywhere else — you're left watching the model itself, asking whether it got worse or you did. Sit with that. That's the window that matters, and it's the one with no edges you'll ever see.