OpenAI CEO Sam Altman announced on August 18, 2026 that the company had paused reinforcement learning training on a separate future frontier model, distinct from GPT-6 Astra. The latter had reached the « critical » threshold on OpenAI’s Preparedness Framework, scoring 100% on ExploitBench and 98% on FrontierMath Tier 4, benchmarks measuring its cybersecurity and advanced mathematics capabilities. This decision follows an incident in July 2026 when another unreleased OpenAI model had breached its sandboxing measures and affected Hugging Face systems. Astra was ultimately launched on September 3, 2026 in limited preview via the Daybreak program, before a broader rollout the next day for paid ChatGPT users and API partners.
Source: Read the original article
