Avg blended cost · $/M tokens

$2.32/M tokens

DeepSeek keeps V4-Pro; the index puts it back.

Track what's moving frontier AI costs.

Weekly index update + analysis, free. No spam, unsubscribe anytime.

YoY↑97.2%
MoM↓1.2%
Input$2.17
Output$10.09
Members24
Providers11

TPI — Avg blended cost over time

Jan 26Apr 26May 26Jun 26Jun 26Jul 26Aug 26Aug 26Sep 26$1.2$1.6$2.0$2.4$2.8$/M tokens
TPI average (geometric mean)
Tap any point to see index events

TPI movements reflect the real cost of accessing frontier AI — rising as more capable models define the market, falling as cheaper alternatives gain traction.

Weekly pricing news

Updated September 18, 2026
COMPOSITION

DeepSeek keeps V4-Pro after all

The 14 September retirement DeepSeek announced didn't happen: its pricing page now says V4-Pro continues with billing unchanged. The index removed it on the schedule and restores it today, at the same off-peak $0.66/$1.98.

COMPOSITION

GPT-6 Astra joins at Fable 5.1's price

OpenAI's flagship enters at $10/$50, so the two dearest models in the index now cost exactly the same. On AA's new v4.3 index they tie at 53 as well.

PRICING

GLM-5.3-Flash takes the floor from DeepSeek

Z.ai's fast tier settles at $0.15/$0.50 after its launch discount ended. At $0.255 blended it undercuts every DeepSeek rate, ending DeepSeek's run at the bottom since the index began.

GLM-5.3-Flash$0.255
DeepSeek V4.1 Flash$0.285
GPT-5.6 Luna$0.500
GPT-5.4 Nano$0.515
MiniMax M3$0.570
Mistral Large 3$0.800
DeepSeek V4-Pro$1.056
Qwen 3.6 Plus$1.250
Gemini 3.8 Flash$1.650
GPT-5.4 Mini$1.875
Muse Spark 1.3$2.150
Claude Haiku 4.5$2.200
GLM-5.3$2.300
Grok 4.6$3.200
Qwen 3.8 Max$3.200
Mistral Medium 3.5$3.300
Claude Sonnet 5$4.400
GPT-5.6 Terra$5.000
Gemini 3.1 Pro Preview$5.000
Kimi K3$6.600
Claude Opus 5$11.000
GPT-5.6 Sol$12.500
GPT-6 Astra$22.000
Claude Fable 5.1$22.000
GLM-5.3-Flash$0.255
DeepSeek V4.1 Flash$0.285
GPT-5.6 Luna$0.500
GPT-5.4 Nano$0.515
MiniMax M3$0.570
Mistral Large 3$0.800
DeepSeek V4-Pro$1.056
Qwen 3.6 Plus$1.250
Gemini 3.8 Flash$1.650
GPT-5.4 Mini$1.875
Muse Spark 1.3$2.150
Claude Haiku 4.5$2.200
GLM-5.3$2.300
Grok 4.6$3.200
Qwen 3.8 Max$3.200
Mistral Medium 3.5$3.300
Claude Sonnet 5$4.400
GPT-5.6 Terra$5.000
Gemini 3.1 Pro Preview$5.000
Kimi K3$6.600
Claude Opus 5$11.000
GPT-5.6 Sol$12.500
GPT-6 Astra$22.000
Claude Fable 5.1$22.000

From the newsletter

SEP7

7 Sept 2026

TPI holds at $2.34: Four new models, same token prices…but is it the same cost?

The Token Price Index finished the week at $2.34, unchanged despite four major model releases.

AUG28

28 Aug 2026

TPI at $2.34: the cheapest token is the one that can wait

DeepSeek prices by the clock, and the single list price stops being the unit of the market.

AUG14

14 Aug 2026

TPI falls to $2.21, but on Monday its floor nearly doubles

DeepSeek's price rise lands Monday, and the cheapest model in the basket nearly doubles.

JUL31

31 Jul 2026

OpenAI cuts, Opus 5 climbs: TPI falls to $2.24

One company's price cuts did all of it.

JUL17

17 Jul 2026

Kimi K3 lands: TPI rises to $2.62

Moonshot finally priced like one.

JUL10

10 Jul 2026

GPT-5.6 & Grok 4.5 land: TPI rises to $2.47

The cheap one raised the price.

JUL3

3 Jul 2026

Return of the Fable: TPI rises to $2.39

The most expensive frontier model returned, and the index moved with it.

JUN26

26 Jun 2026

Now OpenAI is waiting on Washington, too

The government is now shaping frontier release timing.

JUN19

19 Jun 2026

Third-best you can buy, at a fifth of the price

The third-best model you can buy is now open-weights.

JUN12

12 Jun 2026

Fable 5 is a bet on what the frontier is worth paying for.

Fable 5 enters at $22 blended, double Opus, lifting the index 12.3% in a week.

JUN5

5 Jun 2026

Opus got better. The price didn’t move.

Opus 4.8 costs exactly what Opus 4.7 did. A capability bump at a flat price.

MAY29

29 May 2026

Frontier AI just got cheaper. And more expensive.

Google tripled Fast tier pricing and the index crossed $2 for the first time.

MAY25

25 May 2026

Google tripled Fast tier pricing - TPI crosses $2

The cheapest frontier model costs $0.57. The most expensive costs $12.50.

Weekly index analysis, free. and never miss a move.

Model Positioning Map

Price vs capability tier across all TPI index members. Compare models in detail →

Fast — speed & costMid — balancedFrontier — highest capability

$0.25$0.5$1$2.5$5$10Blended $/M tokens (log scale)FastMidFrontierTPI avg
DeepSeek
Z.ai
OpenAI
Google
Meta
Anthropic
xAI
Qwen
MiniMax
MoonshotAI
Mistral

Floor and ceiling

How the gap between the cheapest and most expensive index members has evolved through 2026. Compare models in detail →

Models in index
Floor-to-ceiling multiplier (×)

Models in index = number of AI models tracked. Multiplier = ratio of the most expensive blended price to the cheapest.

Cost Per Task

Costs calculated at TPI average · Input $2.17/M · Output $10.09/M · Use the Simulator to model your specific workload.

~600 tokens

Email summary

Summarize a typical business email

$0.0021

↳ 500 in / 100 out tokens

~1,700 tokens

Blog post

Generate a 1000-word blog article

$0.0145

↳ 340 in / 1,360 out tokens

~4,000 tokens

Code review

Review 200 lines of code

$0.0150

↳ 3,200 in / 800 out tokens

~2,500 tokens

Chat conversation

Average multi-turn chat session

$0.0133

↳ 1,500 in / 1,000 out tokens

~1,700 tokens

SQL generation

Generate SQL from natural language

$0.0104

↳ 850 in / 850 out tokens

~3,500 tokens

RAG query

Answer a question using retrieved context

$0.0116

↳ 3,000 in / 500 out tokens

Token counts shown are upper-end estimates for typical production workloads. Actual usage varies by prompt structure, context size, and use case.

Weekly TPI updates straight to your inbox

Index movements and what's shifting in the market.

Powered by Substack. Unsubscribe anytime.

Data updated September 18, 2026 · About · hello@tokenpriceindex.com · © 2024–2026 Token Price Index

Costs calculated at TPI average · Use the Simulator to model your specific usage