GPT-6 Astra long-context pricing: crossing 272K reprices the entire request
Once GPT-6 Astra input passes 272K, the whole call is repriced at 2x input and 1.5x output, making a 300K request about $8.40; the post suggests compaction and cache reuse.
WHAT THE HELL. CROSS 272K BY 1 TOKEN AND SAM ALTMAN'S OPENAI REPRICES THE WHOLE CALL, NOT JUST THE NEW PART. GPT-6 Astra: 272K in → $4.32 300K in → $8.40 28K more context. Almost double the bill. ━━━━━━━━━━ Below the line: → $10 per 1M input → $50 per 1M output 100K in, 8K out · $1.40 272K in, 32K out · $4.32 ━━━━━━━━━━ Above the line, the whole request flips: → 2x input → 1.5x output The extra 28K isn't what gets repriced. The entire call is. 300K in, 32K out · $8.40 1.05M in, 128K out · $30.60 for one request ━━━━━━━━━━ Where it hurts most: OpenAI dots. Every dot runs on Astra around the clock. Own cloud computer, a browser, 4,000+ apps. Context grows with every task. Leave them working overnight and they drift past 272K while you sleep. ━━━━━━━━━━ The fix is boring. One seat per job. dot works → context hits 272K → compaction pushes old history out → cached input at $1 per 1M instead of $10 → the dot keeps working under the line ━━━━━━━━━━ The model is strong. On OpenAI's own numbers it beats Claude Fable 5.1: → BenchCAD · 95.9 vs 84.3 → Terminal-Bench Science · 64.6 vs 52.6 The bill is where people get hurt. ━━━━━━━━━━ Read before you copy: → prices and scores are from OpenAI's launch material → $8.40 is a modeled call, not your invoice → caching and compaction have their own costs, so run one real night before you trust the math Steal this: → keep the dot's working context under 272K → compact old history → reuse cached input One-person desk. Dots do ~90%, you only say yes. Full price table in the image. Ready to copy. Save this. Follow for the bills nobody reads.
This post quotes another post: x.com/donatelloxpro/status/2109320196649619552