In-depth research from 42 Robots AI on applied AI, model evaluation, and production ML systems.
Eight years of models, GPT-1 to 2026, tested on real production work. We tested 36 AI models on 25 practical task types. About 87% of the measured gain happened by 2022, and on 16 of 25 tasks a model released before 2026 matches or beats the best 2026 model.
Read the article →Methodology and findings from 30 evaluations and 20,110 tests across 9 production models from OpenAI, Google, and Anthropic. Three assumptions about LLM evaluation that the data falsifies.
Read the paper →