Opus 5.5 Outperforms Human Researchers

🔥 What happened Opus 5.5 beats human experts on TasteVal, a new benchmark for experimental research taste. The model matches expert performance with 2.3x less compute and costs only 1/30 per run. 💡 Why it matters Frontier models' compute multipliers now double every 3 months – a huge leap from 14 months before. This means AI progress is accelerating fast, drastically revising forecasts for future models upward. ⚡ Our take If AI has better research taste than humans, who needs us? Experts should upskill – or focus on interpreting results.