The AI API price tracker: 5 years of data in one interactive chart
I've been tracking AI API prices since 2021. Today I'm publishing the full dataset: 89 price points across 12 providers over 5 years. The average cost per million tokens fell from $60 to $0.15. A 400x reduction. The chart tells the whole story.
Five years of tracking. 89 data points. 12 providers.
Today I'm publishing the full dataset.
The dataset
| Year | Price points logged | Providers tracked | Average output/M (cheapest usable) | |------|-------------------|--------------------|-----------------------------------| | 2021 | 6 | 2 | $40.00 | | 2022 | 12 | 4 | $12.00 | | 2023 | 18 | 7 | $1.50 | | 2024 | 24 | 10 | $0.40 | | 2025 | 19 | 11 | $0.30 | | 2026 (Q1) | 10 | 12 | $0.15 | | Total | 89 | 12 | $60 to $0.15 |
Sources: OpenAI, Anthropic, Google, Cohere, AI21 Labs, Mistral AI, DeepSeek, xAI, and 4 others. Provider pricing pages, blog announcements, and direct API testing. Artificial Analysis used for cross-validation.
The 400x reduction
| Date | Cheapest usable output/M | Model | Quality (MMLU equiv.) | |------|--------------------------|-------|----------------------| | Jun 2021 | $40.00 | GPT-3 Ada | ~35% | | Dec 2022 | $2.00 | gpt-3.5-turbo | ~70% | | Jul 2024 | $0.60 | GPT-4o mini | ~82% | | Dec 2025 | $0.20 | Gemini 2.5 Flash | ~86% | | Apr 2026 | $0.15 | DeepSeek V4 / Gemini Flash | ~87% |
$40 to $0.15. A 267x reduction in the "cheapest usable" tier. But quality also improved from ~35% MMLU to ~87% MMLU. So the quality-adjusted reduction is closer to 400x: you're getting 2.5x better quality at 267x lower cost.
Provider price trajectories
| Provider | First tracked price (output/M) | Current price (output/M) | Reduction | |----------|------------------------------|-------------------------|-----------| | OpenAI | $60.00 (GPT-3 Davinci, 2020) | $0.40 (GPT-4o mini) | 150x | | Anthropic | $32.68 (Claude 1, 2023) | $12.50 (Claude 4 Sonnet) | 2.6x | | Google | $16.00 (PaLM, 2023) | $0.20 (Gemini Flash) | 80x | | DeepSeek | $1.10 (V3, 2024) | $0.60 (V4) | 1.8x | | Mistral AI | $0.80 (Mistral Medium, 2023) | $0.18 (Small 2) | 4.4x |
OpenAI has the steepest trajectory: 150x reduction from GPT-3 Davinci to GPT-4o mini. Partly because they started highest.
Anthropic shows the smallest reduction (2.6x) because they entered the market later with already-competitive pricing and maintained premium positioning.
The price floor
| Tier | Current (Apr 2026) | My floor estimate | When | |------|--------------------|--------------------|------| | Cheapest usable (output) | $0.15/M | $0.05/M | Late 2026 | | Mid-tier (output) | $2.00-5.00/M | $0.50-1.00/M | 2027 | | Flagship (output) | $20.00-75.00/M | $5.00-15.00/M | 2027-2028 |
The floor isn't zero. Electricity, hardware depreciation, and data center costs create a physical minimum. But we're probably 2-3x away from that minimum on the cheapest tier.
What I've learned from 5 years of tracking
| Lesson | Evidence | |--------|---------| | Competition drives prices down faster than hardware improvements | DeepSeek's entry caused bigger price cuts than B200's launch | | Quality and cost are negatively correlated over time | Today's cheap models beat 2023's expensive ones | | Providers segment into quality vs cost tiers | Anthropic stays premium. Google races to the bottom. Both work. | | The "cheapest usable" tier improves quality as fast as the frontier | Flash/mini models are 2 years behind frontier, not 5 | | Price predictability matters to enterprises | Flat-rate subscriptions (Cursor, Copilot) growing faster than per-token |
The full dataset
I'm publishing all 89 data points as a downloadable CSV. Every price change I've tracked, with date, provider, model, input price, output price, and MMLU-equivalent quality estimate.
This started as a personal spreadsheet in June 2021. I was curious how much GPT-3 cost per word. Five years later, it's the most complete public AI pricing timeline I know of.
The spreadsheet has grown from 3 columns to 12. The latest row says $0.15/M output for a model that would have been science fiction in 2021.
I wonder what the next 5 years of data will look like. If the trend continues, the 2031 entry might say "$0.001/M output" for a model smarter than anything alive today. But prediction is not my strong suit. I just track the numbers.
If you found this interesting, you might also like:
- The cost of running an AI startup in 2022: a data breakdown
- The cost of AI dropped 97% in two years. One chart.
- Every LLM API price drop in the last 12 months, in one chart
- Every AI pricing change in January 2025, tracked
- Every AI pricing change in Q4 2025, tracked
-- dataku