AI API Token Price Decay 2022-2026: 12 Frontier Models Tracked
Per-model time-series of input and output token prices across 32 priced versions of 12 frontier model families (GPT-3.5 through GPT-5; Claude 1 through Claude 4; Gemini 1.0 through 2.0 Flash; Llama 2/3/3.1 hosted; DeepSeek V2/V3; Mistral Large 2). Cost per a 1k-input/500-output reference task fell roughly 1,000x on the frontier-mini tier between Mar 2023 and Feb 2025; Claude Opus held $15/$75 per million for 26 months across Opus 3 and Opus 4; open-weight hosted pricing leads each closed-weight cost cut by ~73 days median.