Applied AI on open model

Anthropic and OpenAI are booming, but so are providers of open-source AI models and other cheaper alternatives, thanks to businesses using open-source to control their costs, we reported in detail this morning. But there’s more.

Together AI, the neocloud which rents out Nvidia cloud servers and access to open source models, raised its annualrevenue projections at least three times in the past few months due to the growing usage of open source models, CEO Vipul Ved Prakash told us. (Together was generating around $1 billion in annualized revenue as of March.)

The number of tokens processed through its cloud soared to 400 trillion this month from just 30 billion tokens at the same time last year, with much of the growth occurring in the past six months, he said. If you were spending those tokens on DeepSeek v4 Pro, which roughly costs 18 cents per million tokens, that increase would mean the difference of about $70 million in monthly spending. (An AI agent creating a simple document might consume anywhere from 1,000 to 15,000 tokens).

Prakash believes open source models will eventually account for the vast majority of AI usage globally.

At Hugging Face, which operates a repository of open source AI models, paying subscribers to its repository doubled between January and June, co-founder and CEO Clem Delangue told us, without providing specifics.

His co-founder Thomas Wolf said he’s seen a “huge sober, waking-up in companies” concerned about AI costs and getting locked in to using specific AI vendors. 

For their part, OpenAI and Anthropic have said that new models’ prices are more than justified, given how much more they can accomplish for customers. It’s worth pointing out that open source models aren’t a monolith, and figuring out their relative strengths and weaknesses can take time.

The boom in cheaper models extends to providers of so-called model routers that allow companies to easily switch between AI models for different tasks to save on cost. Major firms such as Cisco and Adobe use such routers to temper costs.

“We’ve seen a huge surge in demand over the last six months,” said Not Diamond CEO Tomás Hernando Kofman of the startup’s router. The router allows companies doing AI coding tasks to shift to older and cheaper Anthropic models, which he says saves them 20% to 40% in costs compared to using Anthropic’s most expensive AI.

Providers of such routers and “harnesses,” software that helps an AI agent connect to applications so it can perform complex white-collar work more efficiently, also include Martian AI, Factory, Portkey, OpenRouter, CodeStrap and Vercel.