AI Price Index

Monthly index of AI model API list prices · January 2026 = 100 · latest: October 2026

AI API list prices are not falling. In October 2026, the median chat model cost $1.50 per million tokens on its maker's own API, +23.7% compared with January 2026. Leaving out any single lab, the change ranges from −0.6% (without Alibaba (Qwen)) to +27.6% (without MiniMax), so read the direction, not the decimal.

index123.70% vs last month
median price$1.50per 1M tokens, blended
budget / premium$0.20 / $9.2010th / 90th percentile
tracked175models from 12 labs

// the index, january 2026 = 100

8090100110120130Nov 2025Jan 2026Mar 2026May 2026Jul 2026Sep 2026
median (the index)premium, 90th percentilebudget, 10th percentileJanuary 2026 = 100

// what is behind the number

Labs rarely change the price of a model. 76 models were on sale in both January 2026 and October 2026; 1 of them has a different list price today. Across the 12 months of the index we recorded 15 list-price changes (10 cuts, 5 increases), or about 0.8% of models in a given month.

The catalog moves the median. Since January 2026, 99 models were added at a median of $1.5625 and 72 were retired at a median of $1.525. The 76 that stayed have a median of $1.0625.

Both ends of the market moved. The budget end (10th percentile) went from $0.175 to $0.20, the premium end (90th percentile) from $8.25 to $9.20.

// month by month

monthindexchangemedianbudgetpremiummodelsaddedretiredrepriced
October 2026123.70%$1.50$0.20$9.2017514164
September 2026123.7+20%$1.50$0.218$8.8017714102
August 2026103.1+2%$1.25$0.186$7.201731082
July 2026101−9.3%$1.225$0.175$10.0017111170
June 2026111.30%$1.35$0.175$10.00177830
May 2026111.3+4.9%$1.35$0.175$10.001728245
April 2026106.2+5.1%$1.2875$0.175$10.001881502
March 20261010%$1.225$0.175$7.501731300
February 2026101+1%$1.225$0.175$7.751601200
January 2026100−1%$1.2125$0.175$8.25148300
December 20251010%$1.225$0.175$9.001451200
November 2025101—$1.225$0.175$9.50133000

// median launch price by quarter

quartermedianmodels
2025 Q4$1.12525
2026 Q1$1.4530
2026 Q2$2.0028
2026 Q3$2.1531
2026 Q4 (so far)$1.5258

Models by release date, at the first price we saw. Cohorts are small and swing with a single lab's launches: treat this as a direction, not a measurement.

// cheapest model by capability

capabilityJanuary 2026October 2026
tool calling, 128K+ context$0.0656$0.055 Qwen3.7 Flash
reasoning and tool calling$0.0875$0.055 Qwen3.7 Flash
1M+ context$0.0656$0.055 Qwen3.7 Flash
image input$0.0656$0.055 Qwen3.7 Flash

Blended list price on labs' own APIs, models still on sale. Inference hosts can be cheaper: see hosts.

// median price by lab, january 2026 vs october 2026

labJanuary 2026October 2026changemodels now
Anthropic$6.00$9.00+50%14
Cohere$4.375$4.3750%10
OpenAI$3.50$4.1875+19.6%38
Moonshot AI$1.1375$2.8563+151.1%4
Meta—$2.00new4
xAI$2.725$1.5625−42.7%8
Google$0.2625$1.50+471.4%15
Z.AI (Zhipu)$0.95$1.00+5.3%14
Alibaba (Qwen)$0.875$0.8750%43
DeepSeek$0.315$0.6263+98.8%2
MiniMax$0.525$0.5250%8
Mistral$0.75$0.2625−65%15

A lab's median also moves when it adds or retires models, not only when it changes a price.

// methodology

What it measures. The median blended list price of every paid chat model that 12 labs sell on their own APIs, at the end of each month. Blended means (3 × input + output) / 4, in USD per million tokens. Each model counts once.

What it leaves out. Inference hosts (their past prices are not reliably known), free tiers and promotions, batch, caching and volume discounts, long-context surcharges, and embedding, image, audio and realtime models.

What it cannot tell you. It is not weighted by usage, because labs do not publish usage. It is not adjusted for quality, so it says nothing about the price of a given level of capability. The median also jumps between common price points, so a single month can move more than the underlying market did.

Where the data comes from. models.dev, a community-maintained catalog (MIT), which we cross-check daily against pydantic's genai-prices. Months up to October 2026 were rebuilt from the models.dev git history, catalog by catalog, so models retired since then still count in the months they were on sale. Running that rebuild on the live data reproduced every price, context window and capability across all 175 models. Later months are captured at month end.

Published months do not change. Each month is frozen on the first day of the next one (November 2026 closes on 2026-12-01). Corrections are logged on the about page.

// use this data

series.csvseries.json2026-10.json (every model)

Free to reuse under CC BY 4.0, with a link to this page.

cite it

AI Signals. "AI Price Index, October 2026." https://aisignals.dev/price-index/2026-10

embed the chart

<a href="https://aisignals.dev/price-index"><img src="https://aisignals.dev/price-index/chart.svg" width="640" height="320" alt="AI Price Index: the median AI model costs $1.50 per 1M tokens"></a>

// questions

Are LLM API prices going down?

AI API list prices are not falling. The median model on labs' own APIs cost $1.2125 per million tokens in January 2026 and $1.50 in October 2026. Leaving out any single lab, the change ranges from −0.6% (without Alibaba (Qwen)) to +27.6% (without MiniMax), so read the direction, not the decimal. This index tracks list prices, not what a given level of capability costs, and the two can move in opposite directions.

Does a higher index mean AI got more expensive to use?

Not necessarily. The index is not adjusted for quality or for usage: a model at the same price can be far more capable than one from a year earlier, and most teams can switch to a cheaper model. It measures what labs charge, not what you end up paying for a given result.

What is the blended price?

A single price per million tokens that weights input and output 3:1, (3 × input + output) / 4, a typical mix for chat and RAG workloads.

How often is the AI Price Index updated?

Once a month. Each month is frozen on the first day of the following month and is not revised afterwards; any correction is logged on this site's about page.

Can I use this data?

Yes. The series and the monthly snapshots are free to reuse under CC BY 4.0, with a link to this page as the attribution.