The inference cost collapse, in one chart

I've been tracking AI API pricing since 2021. Five years of spreadsheets. And when I finally plotted the whole thing on one chart, I had to double check the numbers.

They're correct. They're just hard to believe.

The table

Here's what it costs to process 1 million input tokens across the major frontier models, sorted by release date:

| Model | Provider | Release | Input $/1M | Output $/1M | |-------|----------|---------|-----------|-------------| | GPT-4 | OpenAI | Mar 2023 | $30.00 | $60.00 | | Claude 2 | Anthropic | Jul 2023 | $8.00 | $24.00 | | GPT-4 Turbo | OpenAI | Nov 2023 | $10.00 | $30.00 | | Claude 3 Opus | Anthropic | Mar 2024 | $15.00 | $75.00 | | GPT-4o | OpenAI | May 2024 | $2.50 | $10.00 | | Claude 3.5 Sonnet | Anthropic | Jun 2024 | $3.00 | $15.00 | | GPT-4o mini | OpenAI | Jul 2024 | $0.15 | $0.60 | | Claude Haiku 4 | Anthropic | Dec 2024 | $0.25 | $1.25 | | Gemini 2.5 Flash | Google | Feb 2025 | $0.15 | $0.60 | | DeepSeek R1 | DeepSeek | Jan 2025 | $0.55 | $2.19 | | GPT-5 mini | OpenAI | Aug 2025 | $0.40 | $1.60 | | Claude Sonnet 4.6 | Anthropic | Mar 2026 | $3.00 | $15.00 | | Claude Opus 4.6 | Anthropic | Mar 2026 | $15.00 | $75.00 |

The Opus line hasn't moved. That's interesting on its own. But look at the mid-tier. GPT-4 cost $30 per million tokens in March 2023. GPT-4o mini costs $0.15 for roughly similar capability on simple tasks. That's a 200x reduction in 15 months.

What this means

The frontier is getting cheaper, but the real story is in the mid-tier. The models that are "good enough" for 80% of use cases now cost essentially nothing. A million tokens of Claude Haiku 4 costs less than a cup of coffee.

I keep asking myself: what happens when it drops another 100x? When a million tokens costs $0.001? When running AI becomes as cheap as sending an email?

I don't have the answer. But I have the spreadsheet, and the trend line is almost vertical.

-- dataku