Pricing WatchApril 14, 20253 min read

The cost of AI dropped 97% in two years. One chart.

In March 2023, GPT-4 cost $60 per million output tokens. Today, GPT-4o mini costs $0.60. Same-class quality, 100x cheaper. I made one chart. That's the whole article. Sometimes the data speaks for itself.

Sometimes the data speaks for itself.

| Date | Best available model | Output price per million tokens | Quality (MMLU) | |------|---------------------|-------------------------------|----------------| | Mar 2023 | GPT-4 | $60.00 | 86.4% | | Nov 2023 | GPT-4 Turbo | $30.00 | 86.4% | | May 2024 | GPT-4o | $15.00 | 88.7% | | Jul 2024 | GPT-4o mini | $0.60 | 82.0% | | Dec 2024 | Gemini 2.0 Flash | $0.40 | 83.5% | | Apr 2025 | Gemini 2.5 Flash | $0.30 | 85.8% |

Sources: OpenAI, Google, Artificial Analysis.

In March 2023, the best model cost $60 per million output tokens.

In April 2025, you can get comparable quality for $0.30 per million output tokens.

That's a 200x reduction in two years. Quality went up. Price went down 200x.

GPT-4o mini at $0.60/M output tokens achieves 82% on MMLU. GPT-4 at $60/M achieved 86.4%. You lose 4 points of MMLU for a 100x price reduction. For most applications, that's an absurdly good trade.

And Gemini 2.5 Flash at $0.30/M gets 85.8% on MMLU. Better than GPT-4o mini, half the price, and nearly matching the original GPT-4 on benchmarks.

What $0.30 per million tokens means

| Task | Tokens | Cost | |------|--------|------| | Generate a 500-word email | ~700 | $0.00021 | | Summarize a 20-page report | ~3,000 | $0.0009 | | Generate 100 product descriptions | ~50,000 | $0.015 | | Process 10,000 customer support tickets | ~5,000,000 | $1.50 |

Ten thousand customer support responses for $1.50.

I don't have anything clever to add. The chart says everything.

The only other technology I know of that deflated this fast was DNA sequencing, which dropped from $100M per genome (2001) to $1,000 (2014). That was 100,000x over 13 years.

AI inference pricing dropped 200x in 2 years. Different starting points, different scales, but the deflation velocity is in the same league.

My spreadsheet has tracked 89 AI pricing data points since 2020. The trend line looks like a cliff. I keep waiting for the floor and it keeps going lower.


If you found this interesting, you might also like:

-- dataku

More from dataku