How dramatic is the drop?
In 2023, processing one million tokens through GPT-4 cost approximately $60. By mid-2026, the same volume through GPT-4o costs under $1. Through DeepSeek, under $0.15. That is a 400x cost reduction in under three years — faster than Moore's Law, faster than cloud storage, faster than almost any technology price decline in recent memory.
What actually drove the collapse?
Three compounding forces — none of which is slowing down:
1. Competition. When DeepSeek R1 launched in January 2025 matching o1's performance at a fraction of the cost, it forced every major lab to cut prices or justify the gap. OpenAI, Anthropic, and Google all reduced pricing significantly within months. Competition is the most powerful cost driver in any market, and the frontier AI market is now genuinely competitive.
2. Algorithmic efficiency. Models are getting smarter per parameter, not just bigger. Techniques like Mixture of Experts (MoE) mean only a fraction of a model's parameters activate per token — dramatically reducing compute per inference. A 2026 model at similar quality to a 2023 model requires far less compute to run.
3. Hardware improvements. NVIDIA's H200 and B100 chips deliver significantly more AI compute per dollar than the H100s that trained most 2023 models. As newer hardware propagates through data centres, inference costs fall independent of any model improvements.
What the cost collapse actually changes
The ROI calculation flipped. In 2023, many AI automation use cases looked marginal on paper — the cost per AI action was close to the cost of the human action it replaced. That calculation has now reversed for most tasks. The question is no longer "can we afford to use AI here?" but "can we afford not to?"
Volume is no longer a barrier. Organisations that process millions of documents, emails, or data records can now automate at a cost that makes economic sense. Customer service, document review, data extraction, content moderation — all became viable at scale in 2025–2026 in ways they weren't in 2023.
Experimentation is free. When an API call costs a fraction of a cent, the cost of trying something is negligible. This has dramatically accelerated the rate of AI adoption — teams can test, fail, and iterate without meaningful budget impact.
If you evaluated an AI use case in 2023 and decided it wasn't cost-effective, that analysis is likely now out of date. Reprice against current API costs — the answer may be different.
Will prices keep falling?
Almost certainly, though the rate of decline will moderate. The main drivers — competition, efficiency improvements, hardware — are all structural and ongoing. The most likely scenario is continued gradual decline rather than another step-change collapse, unless a new efficiency breakthrough (like DeepSeek R1 in early 2025) resets expectations again.