Most AI is getting cheaper. The best models still aren't.
Frontier models, the most capable AI, are getting pricier while commodity AI keeps falling, leaving buyers with a split decision.
Frontier models, the most capable AI, are getting pricier while commodity AI keeps falling, leaving buyers with a split decision.
OpenAI's o1-preview cost $15 per million tokens to read a prompt and $60 to send one back in late 2024. Four months later, DeepSeek R1 listed at 55 cents and $2.19, a 97% discount that repriced the market overnight.
Commodity inference has fallen from about $20 per million tokens in late 2022 to roughly $0.40 today, a 55x decline in less than four years, per Introl's December 2025 unit-economics analysis cited by AI engineer Aman Panjwani in The Register. Frontier pricing is moving the other way: OpenAI doubled GPT-5.5 to $5 input and $30 output per million tokens, and Google priced Gemini Flash 3.5 at a 3–6x premium over its predecessor.
Larridin CTO Ameya Kanitkar told The Register that engineering-ops token costs at his firm rose about 10x between January and July 2026, driven by agentic workloads and Anthropic's shift of corporate customers from per-seat to metered pricing. Larridin's data puts the threshold at 35–40% of a client's AI budget: past that line, more spend does not buy more productivity, and capping there can cut AI costs about 40% with no other change.
Below the line, buy the cheapest model that handles the task. Above it, the real question is whether the work needs a frontier model at all.