The cost of AI is collapsing
When GPT-4 launched in 2023, running it cost a fortune — enough that all but the largest companies couldn't afford to use it at scale. Three years later, equivalent capability costs roughly 95% less. This price collapse is what turned AI from a research demo into something millions of companies actually deploy. The question now is whether the trend continues, plateaus, or reverses as the next generation of frontier models arrives.
Timeline
- March 14, 2023
GPT-4 launches at $30 per million input tokens and $60 per million output tokens — expensive enough that running it across a real product was prohibitive for most companies.
- November 6, 2023
OpenAI announces GPT-4 Turbo at $10 per million input tokens, a three-fold drop on the same generation of capability. The first signal that AI prices fall fast even within a single model family.
- May 13, 2024
OpenAI launches GPT-4o at $2.50 per million input tokens and $10 per million output tokens. The frontier becomes multimodal at roughly half the price of the previous frontier.
- September 12, 2024
OpenAI launches o1-mini, a reasoning model priced significantly lower than the flagship o1, making reasoning capability more accessible to developers.
- January 31, 2025
OpenAI launches o3-mini at $1.10 per million input tokens and $4.40 per million output tokens, bringing frontier-tier reasoning capability below $2 per million tokens for the first time.
- August 1, 2025
Andrew Ng notes that GPT-4o now costs $4 per million tokens (blended rate), down from $36 per million at GPT-4's initial launch in March 2023 — an 89% price reduction over 29 months.
Where things stand right now
The 38-month price compression has been dramatic. GPT-4's launch price of $30/$60 per million tokens has fallen to single-digit prices for equivalent or better capability. This collapse is what enabled AI to move from research demos to mass deployment. The open question is whether prices continue falling or plateau as frontier models become more expensive to train.