Token Pricing: How China and the US Reprice, and the New Tiers

Dark bars are repricing of the same model; light bars are the price gap between a new generation and the one before. Looking at same-model changes alone makes the US vendors look like they only cut, because their increases all come through new generations. From July 2025 to September 2026 the US vendors raised 13 times and cut 19; 12 of the 13 raises came from OpenAI's and Google's mainstream and mid tiers. China raised 12 times and cut 4, a more consistent direction. The three generational cuts on September 22 were the first time the mainstream tier moved down together.

Each line is the launch-month list price of successive generations in the same product tier. US: OpenAI's mainstream line climbed from the GPT-5 low to $11.25, then GPT-6 Sol dropped to $4 on September 22, the first move down since August 2025; Google Flash went from $0.175 to $3 and then halved to $1.5; Anthropic went the other way, cutting Opus to $10 and then $8 while opening a $20 Fable tier above it; xAI cut with each generation to win share. China's three flagship lines only went up: Kimi five-fold, GLM more than double, Qwen up by half before a generational cut. The diamonds are Fable 5 and GPT-6 Astra: both vendors opened a $20 flagship tier above the mainstream line, so the mainstream line dips at the end because the dearest generation was carved out; the flagship tier itself held at $20, and the thesis stands.

Left: the mainstream chain runs from deepseek-chat through V4-Pro; dashed lines are pricing events: ① February 2025 promotion expiry back to list (+173%), ② September V3.1 output up 50% with the night discount removed, ③ December V3.2 cut 64%, ④ June 2026 the V4-Pro generation, ⑤ July double pricing at peak hours, ⑥ August 16 five times at peak and 2.5 times off-peak, ⑦ once September list prices reflected it the blended price went from $0.544 to $1.98, a new high for the chain. Right: Chinese vendors' share of the Vercel gateway after the increase. Tokens rose to 77% and held, spend rose two weeks running to 29%, requests sit at 44%. Volume held, price went up, money came in: the increase came from supply that could not keep up, a price clearing a tight market rather than weak demand. The right chart is a direction gauge, not a market-wide meter: the sample is overseas app developers, excluding direct enterprise contracts and traffic inside China, so read the direction, not the level.

Both charts are direction gauges: Vercel AI Gateway open data, a sample of overseas app developers that excludes direct enterprise contracts and traffic inside China, so read the direction, not the level. Left: DeepSeek is the volume leader; tokens did not fall after the August 16 increase and doubled to over 60% once V4.1 Flash launched on September 10, held through late September, then fell back from October to 29% on a seven-day average as of November 1 (cause not yet checked; DeepSeek volume on OpenRouter kept rising over the same weeks). Because these are shares, everyone else gets squeezed in the chart, which does not mean their volume fell. Right: Anthropic is the money leader, but its share slid from the low 60s to under 40% (35% on a seven-day average as of November 1), with Kimi (K3 at $6 is actually getting paid), Google and OpenAI taking the difference. The money moves line up with US launches and cuts, not with the DeepSeek increase: ChatGPT Work and 5.6 Sol on July 9, the 29% promotional cut on 5.6 Sol plus an 80% cut on Luna on August 21 (OpenAI's share starts climbing from late August), GPT-6 Astra on September 3, and Opus 5.5 down to $8 on September 22. Anthropic lost share, not necessarily dollars; check against Ramp's September spend share.

One row per model; x is the blended price, the bracket is the Artificial Analysis intelligence index. The blue band is the mainstream tier at $3 to $8, the orange band the flagship tier at $10 to $20. As of September 30: the two $20 flagships, Fable 5.1 and Astra, score below Opus 5.5 at $8 and Sonnet 5.5 at $4, the first time the flagship tier is not where the top scores sit; four new generations entered the mainstream tier within a month; DeepSeek V4-Pro at its $1.98 peak price scores only 36.

Left: the premium a fast or peak-hour channel charges over the standard price for the same model, mostly converging around 2x. Right: who started selling a fast channel and when, from xAI's first in April 2025 to Astra shipping with one at launch in September 2026, one vendor to six in 18 months. Token pricing has split into several dimensions: intelligence, speed, time of day and data rights.

Left: one dot per model, colour by launch period, dashed lines are each period's price frontier; on September 30 the top of the frontier sat outside the $20 tier for the first time, with Opus 5.5 buying the highest score at $8. Right: the lowest cost of reaching an intelligence score of 30, 40 or 50; the 40 line fell from $10 to $0.24 in four months, the 50 line from $10 to $4 in two. A falling application-layer bill may just mean a switch to a cheaper model of equal ability, not falling usage. Note that the cheap end of the frontier is mostly Chinese and open models; US enterprises still buy mainly from the big four.

Left: one price-performance frontier per half year; the frontier moving up and to the left means the same ability keeps getting cheaper and the same price buys more ability. Right: the lowest cost of reaching a fixed accuracy threshold, on a log axis, where the downward slope is the speed of price deflation. This is algorithmic and architectural efficiency, and it holds at the same time as the hardware price on the compute board has stopped falling.

Left: x is the actual cost of running the full intelligence-index benchmark (reasoning tokens included); top-left is strong and cheap; dashed lines are the frontier of each period. On a task basis the ranking flips: GPT-6.1 Sol reaches a score 1.7 points below Astra for an eighth of the money. Right: task cost per point of intelligence; the dearest is still Fable 5.1 at top effort, $246 per point, while the cheapest are GPT-6 Luna, MiMo and GLM 5.3 Flash at $2 to $7 per point. Compare the same effort level across models, or the comparison is unfair.

Left: how many tokens each model burns on the same benchmark; Grok 4.7 lists at $3 but talks the most, so its task cost is high, while GPT-6.1 Sol reaches a high score with the fewest tokens. Right: CursorBench 4.0, an agentic coding benchmark; each line is one model swept across effort levels; Opus 5.5 at high effort scores almost the same as its maximum for a third of the money, the new value sweet spot; task cost and list price can differ by 17 times.

Left: blended price of each vendor's flagship relay across generations; steep deflation from 2023 to mid-2025, and from the second half of 2025 new flagship generations stopped getting cheaper. Right: the small-model line is the cost floor for high-volume, low-value tokens and decides whether AI features fit inside free products; small models are rising too, Haiku 4.5 costs four times 3 Haiku, and only OpenAI keeps pushing small-model prices down.

From the LiteLLM monthly price history; the size is the month-on-month change in blended price. Chinese vendors' time-limited promotions and off-peak discounts mostly do not reach list prices, so their actual count of changes is understated.