The price of intelligence
The best offer has climbed ×205 in 75 days: from 15.2 index points per dollar (mixtral-8x22b, released 2024-04) to 3115.2 (mistral-nemo, released 2024-07). The typical new model has not followed: the median 2026 release buys 42.3 points per dollar, against 116.7 for 2025 releases. Newer models score higher (median 153.5 against 139) and charge more (median $3.48 against $1.20 per 1M tokens).
Intelligence per dollar, by release date
index points per $ · log scale · click a lab to isolate it, a dot for its page
What a level of intelligence costs, week by week
Every Monday, the lowest price (input + output per million tokens) of any offer whose model scores at least 150, then 160, on the Epoch Capabilities Index. The line drops when a cheaper model reaches the level, or when a host cuts its price.
Score: Epoch Capabilities Index by Epoch AI (CC BY 4.0), today's score for every model · prices: each offer's price as published on that date, from our price history, which starts on 5 August 2026.
Year by year
The record is one model. The median is what a typical release of that year buys per dollar. A year with fewer than 5 scored models is shown but not compared.
| 2026 | 42.3 | qwen3.7-flash 900 pts/$ |
| 2025 | 116.7 | gemma-3-4b 1288.9 pts/$ |
| 2024 | 162.8 | mistral-nemo 3115.2 pts/$ |
Intelligence
The Epoch Capabilities Index by Epoch AI, published under CC BY 4.0: one score per model, combined from dozens of benchmarks. Higher is better; the scale has no zero, so a ratio to price compares models with each other, not to an absolute. We report it, we do not produce it.
Price
The cheapest live blended offer (input + output per 1M tokens) across every host of the model; $0 free tiers excluded.
What is missing
Retired models are not priced any more and do not appear, so older years show only the models that survived. Price by model age
See also the current intelligence-per-dollar ranking, the price × intelligence frontier and the methodology.
FAQ
How fast is the price of intelligence falling?
The best offer has climbed ×205 in 75 days: from 15.2 index points per dollar (mixtral-8x22b, released 2024-04) to 3115.2 (mistral-nemo, released 2024-07). The typical new model has not followed: the median 2026 release buys 42.3 points per dollar, against 116.7 for 2025 releases. Newer models score higher (median 153.5 against 139) and charge more (median $3.48 against $1.20 per 1M tokens).
What does 'intelligence per dollar' mean?
Each model's Epoch Capabilities Index score (Epoch AI, combined from dozens of benchmarks, CC BY 4.0) divided by its cheapest blended API price per 1M tokens. It answers: how much measured capability does $1 of tokens buy?
Why do the newest flagships sit low on this chart?
Because peak capability carries a price premium. The model with the highest score today, GPT-6 Astra (OpenAI), costs $60.00 per 1M tokens; the first record holder on this chart cost $8.00. A flagship can score highest and still buy the least intelligence per dollar.