The price of AI is collapsing. The frontier is getting pricier.
Cost
There are two movements in the price of AI right now, and they run in opposite directions. Confusing them is expensive.
Movement one: a given capability becomes almost free
Stanford's AI Index measured what it cost to run a model scoring at GPT-3.5 level on A widely used benchmark of multiple choice questions across 57 subjects, from law to physics.. In November 2022 that was 20 dollars per million Text split into the small pieces a model computes on. An English word is typically one to two tokens. Model pricing is always quoted per million tokens.. By October 2024 the same performance cost 0.07 dollars. A drop of more than 280 times in around eighteen months.
| From | Cheapest model available | Price per M tokens |
|---|---|---|
| November 2021 | GPT-3 | $60 |
| March 2023 | GPT-3.5 Turbo | $2.00 |
| November 2023 | GPT-3.5 Turbo | $0.75 |
| December 2023 | Llama 2 7B | $0.20 |
| May 2024 | Gemini 1.5 Flash | $0.13 |
| July 2024 | Llama 3.1 8B | $0.10 |
The cheapest model you could call, at 6 points in time. From 60 dollars to 10 cents in under 3 years, a six hundredth of the price. Source: Epoch AI.
Epoch AI finds the same across six A fixed test that models are run on, so results can be compared across models and over time.: declines between 9 and 900 times a year, median around 50. They caveat that the fastest drops are recent and may not persist.
What you pay most for today is probably the cheapest thing you buy in eighteen months.
Movement two: the frontier is pulling away
In aggregate prices are sharply down. But break out the last twelve months and the segments diverge: frontier pricing up 36.4 percent, mid-tier down 35.8 percent.
So the gap between the very best and something good enough is widening. The cost of picking the wrong tier is higher today than last year.
The spread is wider than people think
The median across 143 tracked models is 1 dollar per million tokens in and 3.60 out. But between the most and least expensive frontier model the gap is 643x, across models you call with nearly identical requests.
| Model | In | Out | When it is the right call |
|---|---|---|---|
| GPT-5.5 pro | $30 | $180 | Long reasoning where mistakes are expensive |
| Claude Fable 5 | $10 | $50 | The hardest steps in a loop |
| GPT-5.6 Sol | $5 | $30 | Frontier, broadly applicable |
| Claude Opus 5 | $5 | $25 | Frontier, broadly applicable |
| Kimi K3 | $3 | $15 | Upper mid-tier |
| GPT-5.6 Terra | $2 | $12 | The workhorse in most systems |
| Gemini 3.1 Pro | $2 | $12 | Workhorse, long contexts |
| Grok 4.5 | $2 | $6 | Mid-tier at a low price |
| DeepSeek V4 Flash | $0.14 | $0.28 | Classification, extraction, routing |
Dollars per million tokens, retrieved 7 August 2026. We cross-checked three independent price trackers and kept the models where at least two agreed. Prices change monthly, so use the table to understand orders of magnitude, not to budget.
Note where the workhorses sit. A system running everything on the top model pays 15 times what one running mostly mid-tier pays, on tasks where the quality difference is often not noticeable.
Meanwhile quality is converging
As of March 2026, Anthropic, xAI, Google, OpenAI, Alibaba and DeepSeek all sat within 25 A rating number borrowed from chess. Used to rank models by who wins head to head comparisons.of each other on the Arena leaderboard. Competition has moved from capability to price and reliability. Yet capability moved a great deal too: on Humanity's Last Exam the best result went from 8.8 percent in 2025 to 38.3 percent by April 2026.
What this means for you
Do not build the system around one model
An architecture locked to one vendor gets expensive on its own when prices move this fast. Put the model choice behind a layer you can swap in an afternoon.
Review anything older than a year
A system built on last year's frontier is probably doing a job the mid-tier now handles at a fraction of the cost. Switching may not be worth it, but the arithmetic is.
Pay for the frontier where it is genuinely needed
This is not an argument for always choosing the cheapest, but for knowing which few steps require the best.
Without evaluation you cannot take the gain
If you cannot Running the system on a set of known examples and measuring how often the answer is right, so you can see whether a change made it better or worse.whether quality held after a switch, the price collapse is something you read about. That is the most common reason companies keep paying last year's price.
Sources
- Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks
- Stanford HAI, AI Index 2025: State of AI in 10 Charts
- Stanford HAI, The 2026 AI Index Report
- IEEE Spectrum, Stanford's AI Index for 2026 Shows the State of AI
- BenchLM, LLM API Pricing Comparison
- CostGoat, LLM API Pricing Comparison
- LLM Stats, AI Trends
- Developers Digest, Frontier Model API Pricing, July 2026
Figures retrieved in August 2026. Prices and benchmark results move quickly, so check the sources before using any single number in a decision.