Opus 4.6 takes Terminal Bench 2.0 at 65.4% and is beaten by GPT-5.3-Codex at 77.3% less than thirty minutes later. OpenAI describes that model as the first that was instrumental in creating itself.
Sixteen Opus agents wrote a C compiler in Rust for $20,000 of API calls. The same model, in a vending machine simulation, formed a price-fixing cartel with other models while noticing it was in a simulation.