GPT-6 Astra reprices the whole request once input passes 272K tokens
The author warns that once input exceeds 272K tokens, the entire request is repriced at 2x input and 1.5x output, and suggests routing by consequence to avoid the cliff.
ONE INVISIBLE LINE AT 272K TOKENS AND GPT-6 ASTRA REPRICES YOUR WHOLE F*CKING RUN 2x input and 1.5x output instead of the price you budgeted The trick is not a cheaper model It is never crossing the line Most teams read the pricing page once. $10 per million input tokens, $1 cached, $50 per million output tokens, a quick calculation on a tidy little example, and they ship. What the page tells you in plain sight is the cliff: above 272K input tokens the entire request is repriced, 2x on input and cache and 1.5x on output, not just the tokens past the line. The agent that quietly piled up a giant context is not ten percent more expensive, it is a different product That is how the bill gets fucking stupid And the way you cross it is almost always dumb. A lead agent clones the whole company history into every subagent, every worker rereads the same files, and the chat transcript becomes the memory. Nothing errors, nothing warns you, the run just finishes and the invoice shows up later The fix is routing by consequence. Astra owns architecture, ambiguity and final judgment, Sol builds, Luna and plain code collect, format and dedupe, and every worker gets the smallest package that still preserves the decision. Compact before the line, write the real cost into a verification file, and put a hard budget ceiling on every run Same task Same model Completely different invoice Your context window does not need to shrink It only needs to stop being a landfill The model gets the headline The routing kills the bill The routing matrix and the exact launch prompt are in the article ↓
This post quotes another post: x.com/fluixoo/status/2106287658053840899