LLM Benchmark Watch

OpenAI GPT-5.6 family — Astra, Cyber, congressional investigation

OpenAI GPT-5.6 family — Astra, Cyber, congressional investigation

GPT-5.6 Sol/Terra/Luna dominate frontier. New Sol Ultrafast mode (Cerebras, 750 tok/s). Beijing neurosurgeon used Sol to prove 22-year-old math conjecture autonomously. IBM partnership for enterprise deployment. Vision benchmarks show 73% object counting but high cost. Price war intensifies: OpenAI offers 50% discount on GPT-5.6 via OpenRouter, Luna traffic exceeds Claude Opus 5 + Sonnet 5 combined. House investigation ongoing. CRITICAL: OpenAI paused its largest RL training run after Astra met the Critical cybersecurity threshold, triggered by a Hugging Face security incident. New: OpenAI's first NVIDIA Vera Rubin racks are live, running training stack for next-gen pre-training.

Sources (2)
Updated Aug 22, 2026
OpenAI GPT-5.6 family — Astra, Cyber, congressional investigation - LLM Benchmark Watch | NBot | nbot.ai