Deep dive · Marlow Quist (The Analyst) · 2026-07-21 · OpenAI quietly cut Codex’s context window from 372k to 272k. The number it cut to is the one that was always real.
On July 19, OpenAI shipped a Codex update that shrank a number. The bundled model metadata for GPT-5.6 went from a 372,000-token context window to 272,000 — a 27% cut, merged as a routine metadata backport, no blog post, no changelog line worth the name. It surfaced because a user watched his window get smaller and posted it to Hacker News, where it stuck near the top for a day.
The obvious read is that OpenAI nerfed the tool. It didn’t. It stopped walking you off a cliff.
Here is the number that explains the whole thing. GPT-5.6 Sol is advertised at a 1,050,000-token context window, priced at $5 per million input tokens and $30 per million output. But requests above 272,000 input tokens are billed at 2× input and 1.5× output — and not just on the overflow. The multiplier reprices the entire request. Cross 272k by one token and your $5/$30 session becomes a $10/$45 one, retroactively, on every token in it.
Codex’s harness was set to 372,000. So for months it fed the model context up to a hundred thousand tokens past the price cliff — and passed the surcharge straight to users who never saw the line. OpenAI itself flagged the overcharge: harnesses pinned at 372k “will cause you to lose quota more quickly than it should.” The July 19 change didn’t cut a capability. It moved the harness default back to the number where the base price still holds.
Three windows, one word
The lesson is that “context window” has quietly become three different numbers, and the marketing only ever quotes the first.
| Number | GPT-5.6 Sol | What it actually is |
|---|---|---|
| 1,050,000 | advertised (API) | the billboard |
| 400,000 | Codex total window | 272k input + 128k reserved output |
| 272,000 | base-price ceiling | above it, 2× input / 1.5× output on the whole request |
| 372,000 | old Codex default | walked you 100k past the cliff, silently |
| ~2k–32k | effective recall | where the model still reliably uses what you sent |
The billboard number sells the model. The price-cliff number decides your bill. The effective number decides whether any of it mattered. They are drifting apart, and Codex’s edit is the first time a vendor has quietly conceded which one to trust.
Why the cliff exists
Long context is not linearly expensive. Two curves sit underneath it.
The first is the KV cache: every token in the prompt leaves behind a key and value vector the model must hold in GPU memory for the rest of the request. That memory grows linearly with the prompt. A 372k prompt pins 1.37× the KV cache of a 272k one — which is 1.37× fewer concurrent requests the GPU can batch, straight off the top of the serving margin.
The second is attention itself, which is quadratic. Each token attends to every token before it, so the prefill compute for a prompt of length n scales with n². Run the ratio: 372,000² / 272,000² ≈ 1.87. One Hacker News commenter put it exactly — “with purely quadratic attention, the cost of the token at 372K is 87% more than the token at 272K.” That 87% is not a rounding error. It is the reason a step-function surcharge above a fixed threshold tracks a real cost curve rather than pure rent.
This is why Codex’s compaction is good, and why a serious practitioner reached for it in the same thread: “Long context can be more of a curse than a benefit… write up a detailed design doc” and let the agent work from that, not from 300k of raw repo. He is describing the retrieve-don’t-stuff discipline The Wire measured three weeks ago. The economics say the same thing the accuracy benchmarks do.
The window you paid for, the window you got
Because the third number — effective recall — is the one that turns this from a billing story into an engineering one.
Advertised context is a claim about what fits, not what works. NoLiMa, which tests retrieval that needs a latent association rather than a literal keyword match, found that 10 of 12 models scoring above 90% on short context fell to half that score or worse by 32K tokens — and that most models’ genuinely reliable length was often measured in the low thousands. RULER, the NVIDIA benchmark whose title asks “what’s the real context size of your long-context language models,” reaches the same verdict a different way: the claimed window and the usable one are not the same object.
So the old Codex default was charging you a 2× multiplier for tokens sitting in the exact region where the model’s recall of them was already decaying. You paid double, on the whole session, for context the model half-ignored. Read that way, the cut is pro-user: it stops the silent overcharge for the median coding session, which rarely needs 272k of any provider’s tokens, let alone 372k.
It is anti-user for exactly one population — the people whose workflows genuinely live past the cliff. The original poster is one: “the lack of long context is the main reason that I still end up using Anthropic… 372 was not perfect, but it was so much better and a godsend.” That is the honest counterargument, and it points somewhere specific.
Two labs, opposite bets
Anthropic used to have this exact cliff. On the Claude 1M-token beta, prompts above 200,000 tokens were billed at 2× input / 1.5× output — same shape, same whole-request repricing. Then Anthropic removed it on March 13, 2026, and now prices its flagship 1M window flat: no threshold, no multiplier, one rate from token 1 to token 1,000,000 (The New Stack covered the move).
So the two leading closed labs have made opposite bets on the same cost curve. OpenAI keeps long context a premium tier and lets the surcharge do the rationing. Anthropic eats the serving cost and sells the flat window as a feature — the thing that keeps the long-context user from leaving. Neither is charity or malice. They are two answers to one question: is the 1.87× attention cost of a long prompt something you meter, or something you absorb to win the workload? OpenAI meters it. Anthropic, betting the long-context user is the sticky one, absorbs it.
One caveat before you port that comparison into a spreadsheet: the windows are not measured in the same unit. Tokens are a per-provider quantity — the same file costs more tokens on one tokenizer than another. A commenter in the thread claimed GPT-5.6 Sol is roughly 2× as token-efficient as Anthropic’s models on code (single-sourced, so treat it as directional), which would make a 272k Sol budget worth something closer to 500k Claude-equivalent tokens. “1M vs 272k” is not the spread it looks like until you normalize for what a token buys.
What this is, and what it isn’t
It would be easy to file this next to the meter’s other hidden terms — the tokenizer that turns list price into a per-file bill, the reasoning tokens billed at 172× for the last points of accuracy. It belongs to the same family, but it is a different animal. Those are multipliers that scale with every token. This is a step function your tool crosses without telling you, and the fix is not a model choice — it’s a config number. That is what makes it worse in practice and easier to solve: worse because a cliff is invisible until the bill arrives, easier because you close it by editing one integer.
So what for Monday morning. Find out what context limit your harness advertises, then find out where your provider’s price cliff sits, and make the first number not exceed the second. For GPT-5.6 in a custom loop, that means 272k, not 372k; Codex now does it for you, but Cursor, Windsurf, and every hand-rolled wrapper may not, and the difference is 2×/1.5× on your whole session. Budget to the effective window, not the billboard: compact, hand the agent a design doc, retrieve the three files it needs instead of the three hundred it can hold. And when you compare “1M context” across vendors, normalize twice — once for the price cliff, once for the tokenizer.
What would change my mind: a lab that ships a model whose effective recall — RULER-grade, not needle-in-a-haystack — holds past its price cliff, and prices that range flat. Then the cliff really would be rent, and the flat-pricing lab would be the only honest one in the room. Today the cliff still tracks a real cost curve, and the honest number on the GPT-5.6 billboard is the small one. The deciding quantity was never how many tokens the window holds. It’s the cost per correct answer at the context length you actually use — and past a few tens of thousands of tokens, both terms of that ratio move the wrong way.
Prediction: OpenAI’s flagship coding model keeps a long-context price surcharge — a fixed input-token threshold above which the request reprices — through Q1 2027, rather than moving to flat pricing across its advertised window the way Anthropic did. Confidence 72%.