The 272K Long-Context Cliff, Fully Specified

Every current model with a genuinely large context window publishes some version of this rule, and it works the same way everywhere it applies: cross a published input-token threshold, and the entire request — input and output both — reprices at a higher rate, not just the portion of input past the line.

Why "full request" is the detail that trips people up

Most tiered pricing anyone has encountered elsewhere works like a tax bracket — the higher rate applies only to the amount over the threshold. This rule doesn't. The moment input crosses the line, the whole request is billed at the higher multiplier, on both input and output. Treating it like a bracket will understate the cost of every request that crosses it.

Not every model publishes it — and that matters

The rule is stated explicitly on most of the larger-context models in the current lineup. It's genuinely absent from several smaller-window models, consistent with their windows sitting well under the threshold in the first place — a real gap, not just an unstated formality, since a small window makes the rule practically irrelevant. More notably, at least one model with a meaningfully large window doesn't state the rule either way on its own page, and this site records that honestly as not published rather than assuming it inherits a sibling model's behaviour.

Why this bites agentic workloads specifically

A single prompt rarely nears a threshold this size. A long-running agentic session, accumulating file content and prior turns into its own context turn over turn, can cross it as a side effect of ordinary operation — without anyone deliberately sending a "large" request at any single point.

What the table below shows

Every current model, whether it publishes this rule at all, and where the threshold sits relative to that model's own context window — so the size of the safety margin, where one exists, is visible model by model rather than assumed uniform across the lineup.