GPT-6.2 Astra finds vulnerabilities at ~$12 each, worth ~$3,000 in bounties

JiaweiLiu_ · x · 2026-11-18

GPT-6.2 Astra and Sol mark a significant jump in cyber capabilities. On the public DeepBinary benchmark (finding vulnerabilities in Windows system binaries), GPT-6.2 Astra at max reasoning effort found 342 unique vulnerabilities across 12 targets, averaging $12 in API cost per finding versus an estimated $3,000 bounty value — 2.4x the bounty value of GPT-6 Astra's findings at 31% lower API cost. On the internal ExploitBench (turning known browser vulnerabilities into working exploits), it hit a 53.5% success rate, 70% higher than GPT-6 Astra. An internal SecureSWE benchmark tests whether models write code that is both functional and secure, giving defenders an edge.

Original post →

More from Models

Models channel →