OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work

Looped transformers—where model inputs loop through layers multiple times—are central to Astra and the flagship. Microsoft’s accidental leak of this technique in a now-deleted GPT-6.1 Sol write-up confirmed OpenAI’s private disclosures. The approach cuts costs and boosts performance but risks undermining chain-of-thought audits. OpenAI’s chief scientist, Jakub Pachocki, has warned that safety monitoring is becoming increasingly fragile. The company redirected ~10% of its compute—hundreds of megawatts’ worth—to alignment post-incidents like Hugging Face’s model leaks.