Leaked V4 GA Scores Show Relaxed Safety Guardrails, Need for RL Advances

teortaxesTex · x · 2026-08-13

Scores from what appears to be V4 GA have surfaced. The model performs slightly weaker than K3 overall, but shows significantly fewer restrictions in traditionally sandbagged domains.

However, the performance gap suggests that scaling up is not enough; the developer needs another round of breakthroughs in reinforcement learning (RL) environments to close the gap with frontier models like 0731.

Original post →

Leaked V4 GA Scores Show Relaxed Safety Guardrails, Need for RL Advances

More from Models