Artificial Analysis describes two years of change since its coverage of OpenAI’s o1-preview. As of September 26, 2026, it says all frontier models use reasoning tokens before answering. Its Intelligence Index has changed too: version 1 used four single-turn, exam-style evaluations covering general knowledge, science, math and basic coding. Version 4.3 incorporates 10 difficult evaluations, including long-horizon tasks for AI agents, challenging coding problems and knowledge work.