3 articles tagged with Benchmarks
GPT-5.4 scored 75 percent on OSWorld-V this week, quietly crossing the 72.4 percent human baseline for real-world software workflows. This is the inflection point enterprise AI has been waiting for, and the shift from copilot to coworker is going to be messier than the hype cycle suggests.
A deep analysis of Google's Gemini 3.1 Pro benchmarks, the scores they published vs the ones they left out, pricing dynamics at $2/$12 per million tokens, and what the three-way race between Google, Anthropic, and OpenAI means for enterprise AI strategy in 2026.
Anthropic claims engineers use Claude for 60 percent of their work with 50 percent productivity gains and that Claude has achieved AGI by some definitions. A rigorous METR study shows experienced developers are actually 19 percent slower with AI tools, exposing a 40-point gap between self-reported and measured performance.