HomeDemandPricing Tiers
How China and the US Reprice
How China and the US reprice by generation
DeepSeek Main Models: Blended Price and Pricing Events (USD per Million Tokens, log)#
DeepSeek's main model was repriced up three times and down twice in two years, and from July 2026 charges double at peak hours.Draft, pending review
How to read this chart
The main line runs deepseek-chat from V2 through V3.2 to V4-Pro. Dashed lines are pricing events: the February 2025 promotion ending, the September increase and end of night discounts, the December cut, the June 2026 generation change, double pricing at peak hours from July, and another increase in August.
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-01
US Models: Launch Price by Generation, Same Tier (USD per Million Tokens, log)#
US price increases all come through new generations: OpenAI's workhorse rose from $3.44 at GPT-5 to $20 at Astra, while Anthropic cut prices.Draft, pending review
How to read this chart
Each line is the launch-month list price of successive generations in the same product tier, not same-model repricing. Diamonds mark $20 premium SKUs.
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-02
China Flagship Lines: Launch Price by Generation (USD per Million Tokens, log)#
China's three flagship lines climb with every generation, and Kimi K3 priced itself straight into the US band; Qwen 3.8 is the first generational cut.Draft, pending review
How to read this chart
Same y-range as the US chart for comparison. Kimi K2 to K3 went from $1 to $6; GLM 4.6 to 5.2 from $0.96 to $2.15; Qwen Max 3.8 fell back to $3 at its generation change.
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-03
China vs US Flagship Price Bands (Snapshot Sep 4 2026)#
Two models now live in the $20 flagship band, Fable 5.1 and GPT-6 Astra; the workhorse band sits at $3 to $5.6.Draft, pending review
How to read this chart
One model per row; x is blended price (log), the bracket is the intelligence score. The blue band is the workhorse band at $3 to $5.6, the orange band the flagship band at $10 to $20.
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-04
Paying for speed
"Paying for Speed": What the Fast Lane Costs (Multiple of List)#
Fast lanes generally cost double for roughly double the speed at the same intelligence; token pricing has split into intelligence, speed and time-of-day dimensions.Draft, pending review
How to read this chart
The fast lane's price as a multiple of the standard price for the same model. Only OpenAI publishes the speed gain (2x to 2.5x speed at 2x price, same intelligence).
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-05
"Paying for Speed" Spreads: Who Started Selling a Fast Lane, and When#
Paying for speed started with xAI in April 2025 and spread to OpenAI, Kimi, MiniMax and DeepSeek within about a year.Draft, pending review
How to read this chart
X is when each fast lane first appeared; y is its multiple of the standard price.
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-06
All pricing events
All Pricing Events: Same-Model Repricing Plus Generational Repricing#
Counting only same-model repricing makes the US look cut-only; add generational changes and since July 2025 it is 14 increases against 16 cuts.Draft, pending review
How to read this chart
Dark bars are same-model repricing, light bars are generational repricing (new generation versus the previous one's launch price; new SKUs excluded). From July 2025 to September 2026 the US saw 14 increases and 16 cuts, and 13 of the 14 increases came from OpenAI and Google.
Source: FinSight compilation and estimates · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-07
Detail: Same-Model Repricing on GA Releases#
Every same-model repricing on record: who, when, and by how much.Draft, pending review
| Camp | When | Model | Change |
|---|---|---|---|
| 中國 | 2025/02 | deep eek-chat（V3 促銷到期） | +173% |
| 中國 | 2025/09 | deep eek-chat（V3.1、取消夜間優惠） | +83% |
| 中國 | 2025/12 | deep eek-chat（V3.2） | -64% |
| 中國 | 2025/09 | deep eek-rea oner（V3.1 統一定價） | -9% |
| 中國 | 2025/12 | deep eek-rea oner（V3.2） | -64% |
| 中國 | 2026/09 | deep eek-v4-pro（8/16 漲價，尖峰價口徑） | +264% |
| 中國 | 2026/09 | deep eek-v4-fla h（同上） | +277% |
| 美系 | 2023/11 | claude-2 | -27% |
| 美系 | 2023/11 | claude-in tant-1.2 | -90% |
| 美系 | 2024/07 | gemini-1.5-pro（預覽價轉正式價） | +900% |
| 美系 | 2024/09 | gemini-1.5-fla h | -75% |
| 美系 | 2024/11 | gpt-4o | -42% |
| 美系 | 2025/01 | o1-mini | -63% |
| 美系 | 2025/02 | claude-3-5-haiku | -20% |
| 美系 | 2025/06 | o3 | -80% |
| 美系 | 2025/09 | gpt-3.5-turbo | -54% |
| 美系 | 2026/03 | gemini-1.5-fla h | -57% |
| 美系 | 2026/08 | gpt-5.6-luna | -80% |
| 美系 | 2026/08 | gpt-5.6-terra | -20% |
| 美系 | 2026/09 | gpt-5.6- ol（8/21 起限時 3 個月，美系首見促銷價） | -29% |
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
Standalone page ↗d3-cn-us-mon-08