Xint Finds OpenAI Billing Bug That Made Frontier Inference Nearly Free
dyn___ · x · 2026-10-09
Xint researcher Tim Becker found an accounting logic bug in the OpenAI API, dubbed GhostCompact: when an inline compaction was followed by a tool call, the API billed only for post-compaction tokens while the tool-call response still contained the full compaction item — enabling nearly-free inference from frontier OpenAI models.
OpenAI confirmed the issue and deployed a fix on September 21, 2026; Xint received a $600 bounty. The write-up also contrasts billing for standalone versus inline compaction in the Responses API.
More from Infra
- DGX Spark prices skyrocket as resale markups soar — natesiggard · 2026-10-09
- OpenAI bots hit 160K fetches for nonexistent URLs in a week, sparking RL-run speculation — gaganghotra_ · 2026-10-09
- Modal's LLM Engine Advisor picks engine, model and config for your inference workload — charles_irl · 2026-10-09
- LFM 2.5 5.4B seen as better laptop pick; 8B A1B lags 2.6B dense — Aggravating-Push-207 · 2026-10-09
- Architect Fi launches Liquid Inference, a router where providers bid per-prompt across 700+ models — markjeffrey · 2026-10-09
- Open-source lithos-metal hits 200+ tokens/s/user on Qwen3.8-27B with one M5 Max — JiaZhihao · 2026-10-09