Open weights win the tokens, Anthropic keeps 64% of spend
Open-weight models carried 56% of everything routed through Vercel’s AI Gateway in August, their first monthly majority, but collected only 14 cents of every estimated dollar spent. Anthropic took 64 cents of that dollar, per Vercel’s September report published Thursday.
Tokens measure work, not money. Models from DeepSeek, Moonshot AI and Z.ai cost less to run than the proprietary ones from US frontier labs. In July the open-weight token share was 36%.
Inside Anthropic’s slice, teams slid down its own price list: most Fable 5 users cut back, and more of that work moved to the cheaper Opus 5 than to any other model. “Lab loyalty doesn’t follow brand, it follows model profile,” Vercel’s report authors note.
Google went the other way, its share of token volume on the gateway falling from 30% to 5%.
Up next
Anthropic and OpenAI start shopping in 20–30 MW slices
20–30 MW is what Anthropic has been sounding out across the U.K. and the Nordics, four sources told CNBC, after booking around 460 MW from Nscale in West Virginia. Small blocks reach usable capacity faster.
Related stories
- GLM-5.3 weights land with a hyperscaler clause
- One model cleared all five test categories at 100%
- Revolut moves 2,000 engineers to self-hosted GLM at $0.30/M
- Cat Wu, Claude Code’s head of product, leaves Anthropic
- Anthropic’s Q3 brought $17B in revenue and an operating loss
- Anthropic’s amended S-1 lists $895B in compute commitments
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.