The 272K Cliff Reprices the Whole Request

The mistake

Budgeting a large request by assuming the long-context repricing rule works like a tax bracket — that only the portion of input past the published threshold gets charged at the higher rate, and everything below the line stays at the ordinary price. It doesn't work that way, and a spreadsheet built on that assumption will understate every request that actually crosses the line.

Why this happens

Most people's only mental model for tiered pricing is the bracket kind — income tax, cloud storage tiers, plenty of ordinary SaaS pricing — where crossing a threshold only changes the rate on the marginal amount above it. OpenAI's rule reads similarly at a glance ("prompts over the threshold are priced higher") and it's easy to skim past the one clause that actually matters: the higher rate applies to the full request, input and output both, not just the excess.

Why it matters

The practical effect is a price jump that looks disproportionate to whatever pushed the request over the line — one more file of context, one more turn of accumulated conversation history — because the whole request repriced, not just the increment. This bites agentic sessions hardest, since they accumulate input turn over turn without anyone deliberately deciding to send a "big" request; the threshold gets crossed as a side effect of normal operation, and the rule isn't published for every model, so the same workload can behave differently depending which model is answering it.

The fix

Treat the threshold as a real constraint on agentic loops, not a rounding concern. Check whether the specific model you're running actually publishes this rule before assuming it applies — it's stated for some models and explicitly not stated for others, and "not published" isn't the same as "doesn't happen." For a session that's accumulating context turn by turn, watch the running input size between turns and trim or summarize before it crosses the line, rather than discovering the jump after the bill arrives.

See also

The Context-Window Planner shows exactly where a given prompt sits relative to both a model's context ceiling and this threshold, side by side.