OpenAI released GPT-6 Astra this week, and it is a generational leap.

OpenAI released GPT-6 Astra this week, and it is a generational leap. Most attention is focused on its reasoning variants, which set new records across the board. However, the non-reasoning base model also has a very impressive story. I compared Astra Non-reasoning against top reasoning models released between 2 and 12 months ago using Artificial Analysis benchmark data.

The numbers:

• Artificial Analysis Intelligence Index: Astra Non-reasoning scores 47.8. Claude Opus 4.8 Max Effort scores 46.4, and GPT-5.5 xhigh scores 45.6.
• Artificial Analysis Coding Index: Astra Non-reasoning scores 76.2. GPT-5.5 xhigh scores 74.9, and Claude Opus 4.8 Max Effort scores 74.3.
• TerminalBench v2.1: Astra reaches 89.1%. GPT-5.5 xhigh scored 83.5%, and Claude Opus 4.8 Max Effort scored 80.1%.

Astra achieves this in a single forward pass without emitting reasoning tokens. It beats models that spend thousands of internal tokens thinking through each prompt. Before June 2026, no reasoning model scored higher on these composite benchmarks.

While this is an impressive engineering result, it creates a serious problem for AI alignment. Frontier labs rely heavily on Chain-of-Thought traces during training and safety evaluations. By inspecting reasoning steps, researchers measure whether a model is scheming, reward hacking, or exhibiting deceptive alignment. With non-reasoning base models, that window into internal decision-making disappears. The computation happens entirely in latent representations, leaving labs with no legible trace to verify why a model chose a specific response. As base models match reasoning architectures, our primary methods for measuring and training alignment lose their foundation.