AI token prices are falling. Your AI bill is not.
The short version
Frontier model prices are 84.5 percent below March 2023, and the cheapest tiers dropped about 96 percent. Yet the median price across 196 models today is $0.75 per million input tokens and $3.55 output, and frontier prices actually rose 66.7 percent over the last year. Falling unit prices and rising bills are both true.
What the deflation looks like
The LLM Token Expenditure Index from Silicon Data fell to 97 cents on 1 September 2026, a record low. That index tracks what buyers actually pay across models, not the sticker price of one product.
Longer run, the numbers are brutal. LLM Cost Hub, which tracks 128 models from 17 providers, puts average annual deflation at 60 to 70 percent. Its tier breakdown since Q1 2023: frontier class down 93.8 percent, from $45 to $2.80 per million blended tokens. Mid-tier down 95.7 percent, from $15 to $0.65. Small models down 96 percent, from $2.00 to $0.08.
What things cost today
| Model | Input / 1M | Output / 1M |
|---|---|---|
| GPT-6 Luna | $0.10 | $0.50 |
| Claude Haiku 5.5 | $0.10 | $0.50 |
| Gemini 3.8 Flash | $0.75 | $3.75 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Gemini 4 Argon | $2.00 | $10.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
For scale, Claude Opus 5 launched at $5 and $25. Opus 5.5 arrived at $4 and $20, which is a 20 percent cut inside a single generation.
The part nobody quotes
BenchLM's index tells a different story depending on which slice you read. Frontier prices are 84.5 percent below March 2023, yes. But over the twelve months to October 2026, frontier prices rose 66.7 percent year on year, while mid-tier prices fell 44.5 percent.
That is not a contradiction. The frontier moved up a tier. What counted as the best model in 2025 is now a mid-tier product, and the new top models are priced for the work they do, not to win a price war.
Meanwhile the median across 196 tracked models is $0.75 input and $3.55 output. Open-weight models sit about 75 percent below proprietary ones, at roughly $0.50 blended versus $2.00.
Why your bill still goes up
Three reasons, and they compound.
Agents use more tokens per task. A chat reply might read 2,000 tokens. An agent that plans, calls four tools and retries twice can read 200,000. Ten times cheaper tokens with fifty times the consumption is a fivefold increase.
Cheaper tiers sell out first. When capacity is tight, as it is in 2026, providers steer volume to the tiers with margin. The price list says one thing; the routing says another.
Retries and context cost money. Prompt caching cuts repeat context by 50 to 90 percent, but only if you build for it. Most teams measure price per token and never measure tokens per task.
What to actually track
Stop tracking price per million tokens. Track cost per finished task, and the share of that task served by the small model.
A support answer that costs 3 cents on a small model and 40 cents on a frontier model is a routing decision, not a pricing decision. Teams that route by task complexity typically cut spend by half without touching quality on the long tail.
The table above is the wrong unit of comparison for that work. You need your own logs.
Questions people ask
Are AI prices really falling?
Per token, yes, about 60 to 70 percent a year on average since 2023. Per task, it depends on how much more work each task now does.
Why did frontier prices rise 66.7 percent?
Because the frontier moved. Models that were top-tier a year ago are now mid-tier, and the new top models are priced higher than those older flagships.
What is the cheapest usable model right now?
GPT-6 Luna and Claude Haiku 5.5 both sit at $0.10 input and $0.50 output per million tokens, with Gemini 3.8 Flash at $0.75 and $3.75.
Read next: the four numbers that decide your AI bill, and why data center power demand is now the constraint.
Sources
CNBC on the LLM Token Expenditure Index (1 September 2026); LLM Cost Hub deflation tracker and provider pages (19 October 2026); BenchLM LLM pricing statistics (19 October 2026); Gravity.fast API price history (4 October 2026).
Read from the primary source on 21 October 2026. Numbers change, so check before you commit.