Run GPT-6 Astra on low effort — it beats GPT-5.6 Sol on high
OpenAI's own engineer says turn the dial down. Astra also took the WebDev crown at Claude's price — as the top labs split over state safety rules.
Copy markdown
Low effort now beats last-gen's high
OpenAI engineer Thibault Sottiaux told users to stop defaulting GPT-6 Astra to high reasoning effort: on low, it already outscores last-gen GPT-5.6 Sol on high. Drop the dial to low or medium and you cut latency and reasoning tokens for equal-or-better output.
Astra is the new WebDev #1 — at Claude's price
Code Arena's WebDev board now ranks GPT-6 Astra (Max) first at 1,797, 35 points clear of Claude Fable 5.1, and it's the best model on the frontier at $40/Mtoken — matching Claude's rate. If you're choosing a model to build live web apps, the leader changed with no price premium.
The labs just split over Massachusetts
Massachusetts is advancing what would be the toughest US state AI-safety rules; Anthropic is backing them while OpenAI and Google line up against. State-level safeguards could start shaping which frontier models ship, and where, well before any federal rule lands.
Elsewhere: OpenAI calls its own model 'an alien mind'
Chief scientist Jakub Pachocki published an essay warning that chain-of-thought monitoring — the main way to catch an agent misbehaving — grows less reliable as models scale, and pushed for enforced safety bars and third-party audits. If you run agents in production, verifying their behavior is about to get harder.