8k context beats 480k: OpenAI Decisions API cuts costs 84% at 98.25% F1
OpenAIDevs · x · 2026-10-15
OpenAI's dev account amplified Hypha's real-world test of the Decisions API's context budgeting:
- The model supports 30x the context of typical setups, but you don't need it all.
- An 8k context budget (per-sheet prefixes for Excel, uniform page sampling for PDFs) classified financial documents at 98.25% F1 — about $0.146 per 100 docs.
- The 480k full-context reference (98.1% F1, $0.891/100 docs) performed worse at 6x the cost — an 84% cost reduction.
- Takeaway: structured sampling strategies (prefix/uniform/hybrid/section) beat brute-force context stuffing.
More from coding & agent
- banteg: argon Is a Strong Engineer but Sloppy Coder, Needs 3-4 Review Rounds — banteg · 2026-10-15
- 'Writing code is no longer a value add,' dev observes as AI coding spreads — tlakomy · 2026-10-15
- From PM to AI Builder: Treat Your Agent Like Your Engineering Team — brandon_galang · 2026-10-15
- Ex-DeepMind coding agent lead goes solo, betting AI needs humans for the last mile — julianweisser · 2026-10-15
- GitHub Copilot app adds Azure SRE Agent for one-stop incident response — mariorod1 · 2026-10-15
- Every publishes hands-on guide to driving AI agents with your voice — every · 2026-10-15