According to Epoch AI analysts, the cost of AI services has decreased by several hundred thousand times over the past five years. On average, AI costs drop by nearly 50% every quarter, or 13 times annually. This rate of decline is four times faster than that of DNA sequencing and 18 times faster than that of lithium-ion batteries.
The analysis is based on models achieving at least 81.25% in the GPQA Diamond benchmark. For instance, the OpenAI o3 model released in January 2025 cost $0.30 per question, while the GPT-5.6 Luna model, released 18 months later, achieved similar results at a cost of $0.0004 per question.
The pace of cost reduction is slowing down in most disciplines, dropping from 66% per quarter initially to 32% per quarter after two years. This trend is influenced by rising costs for electricity and computing hardware in 2026, making it increasingly difficult for developers to lower service prices further.
Developers face challenges in maintaining customer loyalty as rapid price drops and performance improvements make it difficult to justify premium pricing for top-tier models. Companies like OpenAI and Anthropic must compete on the Pareto frontier, balancing cost efficiency with model intelligence, while customers weigh these factors against ecosystem integration and trust.
