Timeline: 2019 — May 2025
Real benchmark data from GPT-2 through Qwen3 and Gemma 3.
Models marked (est.) have estimated scores based on contemporaneous evaluations.
Red Line = SOTA Frontier (best performance at each point in time).

Human Baselines (Challenging Targets):
• GSM8K (Reasoning): 95% - Near-perfect on grade-school math
• HumanEval (Coding): 90% - Skilled programmer "puzzle"-coding
• MMLU (Knowledge): 89.8% - Domain experts across 57 subjects