The Price of Intelligence Just Crashed SAN FRANCISCO - Remember when running GPT-4 cost $30 per million tokens in early 2025? Eighteen months later, that number is $0.12. And the floor keeps dropping. Inference costs fell 40% quarter-over-quarter in Q2 2026 - from $0.20 to $0.12 per million tokens - while latency improved 35% to 340ms p50. The number of available models on major API providers hit 240+. We are living through the fastest commoditization of a transformative technology in computing history. ==chart Metric
| Q1 2026 | Q2 2026 | Change Avg cost/M tokens (mid-range) | $0.20 | $0.12 | -40% Latency p50 | 520ms | 340ms | -35% Available models on major APIs | 160 | 240 | +50% Proprietary model share of API traffic | 68% | 52% | -16pp Open source model share of API traffic | 32% | 48% |
