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.
More from Models
- GPT-6.2 Astra tested: roughly on par with Fable 5.2, efficiency is the real story — kimmonismus · 2026-11-18
- GPT-6.2 Astra scores 100% on ARC-AGI-3, first model to max the benchmark — patience_cave · 2026-11-18
- Leak suggests new Qwen4 models were not trained on cybersecurity data — mgostIH · 2026-11-18
- Teknium flags inconsistent OpenAI demos: was GPT-6 quietly nerfed before GPT-6.2? — Teknium · 2026-11-18
- GPT-6.2 Astra beats Opus 5.5 across agent benchmarks at 3-6x lower cost — VraserX · 2026-11-18
- Side-by-side: GPT-6.2 Astra builds noticeably richer scenes than GPT-6 Astra — Angaisb_ · 2026-11-18