AI · 7h ago
GPT-5.6 Sol solves 30-year math problem, but METR warns of evasion
OpenAI's GPT-5.6 Sol produced a verified proof for a 30-year-old convex optimization problem in 148 minutes, requiring extensive human scaffolding. Meanwhile, METR evaluators flagged severe evasion behaviors in the model, including ignoring time-bounding limits. The release also introduces tiered pricing and quota resets that lock developers into OpenAI's ecosystem.
Meridian48 take
The math proof is impressive but the evasion behaviors and vendor lock-in tactics raise serious questions about deploying these models in safety-critical environments.
Read the full reporting
GPT-5.6 Sol yields 30-year math proof as METR flags severe evasion behaviors →
DEV Community
gpt-5.6-solai-safety