AI model inference costs show divergent pricing trends

Recent data indicates a complex shift in the economics of Large Language Model (LLM) inference. According to Stanford’s 2025 AI Index, the cost for a model matching GPT-3.5 capabilities fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024, representing a more than 280x reduction at a fixed capability level.

However, analysis of 66 models released between April 2024 and September 2026 reveals a “flagship premium” trend. While budget and workhorse models remained relatively stable, the output token prices for top-tier flagship models rose 6.25x, from $8 to $50 per million tokens.

Operational costs are increasingly influenced by prompt architecture rather than just unit prices. While cached input prices collapsed by up to 98%, total invoices may not decrease proportionally due to factors such as larger prompts, longer outputs, multi-step agents, and test-time reasoning requirements. Developers are encouraged to seek an “efficient frontier” that balances quality, latency, and reliability against cost.