What It Actually Costs to Run a Big AI Model

Running a frontier AI model uses thousands of dollars worth of NVIDIA H100 GPUs per minute. The cost per query has dropped ~99% from 2022 to 2026 because of better hardware, smaller specialized models, and engineering tricks like quantization. That price drop is why AI features keep getting added to free tiers.

Some examples

• An H100 GPU rents for $2-$4/hour on cloud services; training GPT-4 reportedly cost $100M+ in compute.
• GPT-4 cost $30 per million input tokens at launch (March 2023). GPT-4o-mini in 2026 costs $0.15 per million — a 200x drop.
• DeepSeek's V3 model in late 2024 trained for under $6M, partly by using cleverer methods on cheaper hardware.
• Inference (running the model after training) is cheaper than training but happens billions of times per day across all users.

Try it!

Visit openai.com/api/pricing or anthropic.com/pricing. Compare the same model from a year ago to today. The drop is real and steady. That's the trend that makes 'AI features in everything' possible.