newsfilter
← Interviews
AI Engineer18 min

Post-Training Gemini 4 for Real Engineering Work — Melvin Johnson, Google DeepMind

Melvin Johnson

Conference Presentation, Keynote, Product Demonstration

Google engineers recently accelerated their frontier model development by launching Gemini 4.5 Argon, Carbon, and Hydrogen within a single month to demonstrate a tenfold increase in delivery velocity. These new models achieved top-tier rankings on industry benchmarks like CursorBench and Terminal Bench v4, delivering significant cost advantages and performance gains while reducing fabrication rates to as low as 1.5 percent. This rapid iteration was driven by a rigorous internal deployment strategy leveraging over 100,000 employees and a refined training methodology that prioritizes hard engineering tasks and anti-gaming evaluation protocols.