The frontier model war just took a wild turn. While Gemini 4 Argon is delayed, Google is reportedly ghost-testing a secret "Carbon" checkpoint in its internal Jetski platform. Insiders claim its coding capacity matches Claude Opus 5.5, fueled by recursive self-improvement (RSI) loops.
Meanwhile, Microsoft officially launched Decision-1, a dedicated scoring model that runs 35x faster than GPT-6 Sol and costs 200x less for routing and labeling.
Are smaller, hyper-optimized agents the real future?