이미지: The Decoder
Summary
- Anthropic captured 65.1% of Vercel AI Gateway's July revenue while processing only 30% of total tokens
- Anthropic's average token price is 4.4x the average of other providers, and its top model Fable 5 held second place with a 13.2% revenue share as new teams poured in
- Total gateway token volume grew 59% and spending grew 37%, but average token price fell 13.6%, showing the market is also moving toward lower prices
- 앤스로픽 매출 점유율
- 버셀 AI 게이트웨이 7월 매출의 65.1%
- 앤스로픽 토큰 점유율
- 전체 처리 토큰의 30%
- 토큰당 단가 격차
- 다른 공급사 평균의 4.4배
- Fable 5 매출 점유율
- 13.2%, Opus 4.8에 이어 2위
- Fable 5 신규 고객 비율
- 7월 신규 팀 10곳 중 9곳
- 전체 토큰 처리량 증가
- 59% 증가
- 전체 지출 증가
- 37% 증가
- 평균 토큰 가격 변화
- 13.6% 하락
30% of tokens, 65% of revenue
Vercel's AI Gateway is a service that lets developers call multiple AI model providers' APIs through a single interface. In its July results, Anthropic took 65.1% of total revenue on the gateway. But the actual token volume Anthropic processed came to only 30% of the total. In other words, it earned far more than the volume of work it handled would suggest. The gap comes down to price: Anthropic's average per-token price is 4.4 times the average of other providers.
Even inside Anthropic, this figure was somewhat unexpected. In early August, Anthropic launched its top-tier model, Fable 5, priced at $10 per million input tokens and $50 per million output tokens — roughly twice the price of competing models. As covered in The Most Expensive AI Model Struggled to Win Over Enterprise Wallets, payments data provider Ramp read this as a sign the pricing was too aggressive, noting that Fable 5 accounted for only 6% of Anthropic's total token sales volume and 11.4% of revenue in its first month.
Vercel saw a different picture
Vercel's July gateway data told a different story. Fable 5 climbed to 13.2% of total gateway spending, placing second behind Opus 4.8. More notably, 9 out of 10 teams that used Fable 5 in July were new customers. Vercel interpreted this as a sign that the model continues to attract new users and that willingness to pay for premium models remains strong.
Ramp and Vercel reached opposite conclusions about the same model from the same company — one seeing slow adoption, the other seeing a surge of new users. Given that each company is only looking at one slice of the overall market, this divergence itself suggests the premium AI model market hasn't yet settled into a clear pattern.
The broader market is also moving toward cheaper options
While Anthropic held its prices steady and captured revenue, the gateway's overall metrics pointed in a different direction. Over the course of July, token volume passing through the gateway rose 59%, and total spending rose 37%. But average token price fell 13.6%, indicating that companies are shifting toward lower-tier models. This aligns with recent price cuts across the industry — OpenAI cut both input and output prices for GPT-5.6 Luna by 80%, and Anthropic itself priced Opus 5 at half the cost of Fable 5.
According to the image caption, xAI and Moonshot AI also captured a notable share of open-weight model spending for the first time in July. Even as low-price competition intensifies, Anthropic stands out as the one maintaining high per-token prices while still growing its revenue share.
| Metric | Value |
|---|---|
| Anthropic revenue share | 65.1% |
| Anthropic token share | 30% |
| Anthropic per-token price | 4.4x average |
| Fable 5 revenue share | 13.2% (2nd place) |
| Fable 5 new customer ratio | 90% |
| Overall gateway metrics | Change |
|---|---|
| Token volume | +59% |
| Total spending | +37% |
| Average token price | -13.6% |
Editor's view
These numbers show Anthropic occupying an unusual position in the pricing war. While OpenAI slashed GPT-5.6 Luna prices by 80% and Chinese models pulled down the token price index by 25%, Anthropic actually grew its revenue share on the gateway. This doesn't mean cutting prices is unnecessary to sell — it means there's a distinct segment of the market where products sell even without price cuts. Many of Vercel Gateway's customers are developers and startups rapidly building prototypes, for whom code accuracy and response quality matter more than per-token cost.
It's also worth noting for practitioners that Ramp and Vercel drew opposite conclusions about the same company's same model. Payment card data suggests slow adoption, while API gateway traffic suggests a surge of new users. Both metrics are real — they're just looking through different windows. Payment data shows how much budget companies have allocated, while gateway data shows which models developers actually call in their code. There can be a gap of several months between enterprise budget approval and developers' actual choices.
There's one practical takeaway domestic teams should draw from this data: choosing a model based on per-token price alone only tells half the story. The fact that a model priced twice as high, like Fable 5, can still attract 90% new customers shows that output quality can override budget decisions. Conversely, in a market where average prices are falling 13.6%, running pilots first with cheaper models and reserving premium models only for cases where results fall short is the more cost-effective strategy. Whether Anthropic's revenue share climbs further in next month's Vercel report, or whether cheaper open-weight models like xAI's and Moonshot's eat into that share, will be the next signal in this pricing war.
