Where the 430× goes
Token prices fell 430× since 2021. Token consumption climbed as AI moved from chatbots to agents. How much of the drop reaches the bill depends on the workload.
Inference fell 430× in five years. Open-weight models have set a floor no company can raise. But how those systems are used has changed. A single chatbot session consumed roughly 1,000 tokens in 2021—it still does. An AI Copilot running in the background consumes around 5,000 tokens daily. A resident agent, Goldman’s term for software that works continuously on your behalf instead of waiting to be prompted, burns over 100,000. Price fell 430×; a resident agent uses 100× more tokens than the chatbot baseline. Net savings: 4.3×. A task that cost six cents in 2021 costs roughly a penny today as a resident agent run. But 4.3× is not 430×, and the difference matters when you’re building a business around one number or the other.
Methodology & sources
GPT-3 Davinci price (Nov 2021): $60/M tokens (OpenAI pricing). DeepSeek V3 (Apr 2026): $0.14/M tokens. Baseline: 1,000-token chatbot session in 2021 = $0.06. Net savings = price drop ratio ÷ token growth ratio. All three token consumption figures are from Goldman Sachs “Decoding the Agentic Economy” (May 2026): chatbot ~1,000 tokens per session; AI Copilot ~5,000 tokens per day; resident agent ~100,000 tokens per day. Source: Goldman Sachs.