AI price index · updated October 12, 2026
How fast is AI getting cheaper?
Key figures
- 97%less for GPT-4o-level capability today than at its launch.
- 36×cheaper: the same capability as GPT-4o for a fraction of its launch price.
- 2.8×the price of today's top-scoring model, Claude Fable 5.1 ($22.00 per million tokens), against GPT-4o at launch.
Launch price against today's price for the same capability
Every model in our table that has been on sale for six months or more and whose launch price we know.
- Claude 3 OpusMar 2024−99%
- Claude 3 HaikuMar 2024−60%
- GPT-4oMay 2024−97%
- Gemini 1.5 FlashMay 2024−61%
- Gemini 1.5 ProMay 2024−96%
- Claude 3.5 SonnetJun 2024−97%
- GPT-4o MiniJul 2024−23%
- Mistral Large 2Jul 2024−95%
- Claude 3.5 HaikuNov 2024−88%
- DeepSeek R1Jan 2025−79%
- o3-miniJan 2025−89%
- Gemini 2.0 ProFeb 2025−91%
- Claude 3.7 SonnetFeb 2025−97%
- GPT-4.1Apr 2025−94%
- Claude Haiku 4.5Oct 2025−90%
- DeepSeek V3.2Dec 2025−32%
- GPT-5.2Dec 2025−19%
- Claude Opus 4.6Feb 2026−60%
- Claude Sonnet 4.6Feb 2026−97%
- Gemini 3.1 ProFeb 2026−96%
The pairs
Blended price per million tokens: 70% input, 30% output, standard list price on the provider's own API.
Nothing on sale today is both at least as capable and cheaper than these were at launch: DeepSeek V3, Gemini 2.0 Flash, and GPT-5 Nano.
The frontier went the other way
Capability got cheap; the best model did not. The highest-scoring model on sale today, Claude Fable 5.1, costs $22.00 per million tokens, more than GPT-4o cost at launch ($8.00). What falls is the price of last year's frontier, not the price of the frontier.
How it is calculated
For every model that has been on sale for at least six months and whose price history starts in its launch month, we take its launch list price. We then find the cheapest model on sale today whose text score is at least as high and compare the two prices.
The text score is the average of iOPTERA's editorial coding, reasoning, extraction and creative scores (0–100). These are composites we set against public benchmark results, not benchmark results themselves; image input is left out so that text-only models can qualify.
Prices are list prices on each provider's own API, standard tier, shortest-context band, blended at 70% input and 30% output tokens. The index updates itself whenever a price or a model in the table changes.
Use the data
The pairs are open data under CC BY 4.0: use them with a link to this page.
Cite it
iOPTERA (2026). AI price index: GPT-4o-level AI costs 97% less than at launch. https://ioptera.com/en/price-index
Embed the chart
Paste this into any page. The chart updates itself, and the link under it credits the source.
<iframe src="https://ioptera.com/embed/en/price-index" title="AI price index" width="100%" height="1074" style="border:0;max-width:720px" loading="lazy"></iframe>
<p style="font-size:13px">Source: <a href="https://ioptera.com/en/price-index">AI price index by iOPTERA</a></p>Link to it
<a href="https://ioptera.com/en/price-index">AI price index by iOPTERA</a>