Artificial intelligence startup Anthropic has overhauled its smallest AI model, Haiku, as companies increasingly opt for AI systems that are nearly as capable as the most advanced tools – but come at a fraction of the cost.
Anthropic said on Wednesday that the updated model, called Haiku 4.5, is about one-third the cost of Sonnet 4, one of its medium-sized models, and one-fifteenth the cost of its most advanced offering, Opus. The new Haiku model performs as well or better than Sonnet 4 on a range of tasks, including coding.
Traditional companies outside of Silicon Valley are more likely to use AI if they have cheaper, yet still capable models to try, Mike Krieger, the company’s chief product officer, said in an interview. It also makes it easier for firms to add more AI to their internal systems that might be used by hundreds or thousands of employees, he added.
“Often, there’s a lot of scale to that,” Krieger said. “Small models really help because they can be a more economical way of deploying into that.”
The vast majority – about 80 per cent – of the company’s revenue comes from companies, an Anthropic spokesperson said, adding that it has more than 300,000 enterprise customers that use Anthropic tools internally, for products or both.
Anthropic’s annual revenue run rate – a calculation of annual revenue extrapolated from the current sales pace – is almost $7 billion, the spokesperson said.
Anthropic also draws revenue from people who pay to use Anthropic’s chatbot, Claude.
San Francisco-based Anthropic, last valued at $183 billion, was formed in 2021 by a group of former OpenAI employees. Its models go toe-to-toe with those of its more famous counterpart.
This is the first time in about a year that Anthropic has updated its smallest model, Krieger said.
In the earliest days of the modern AI boom, executives at companies like Anthropic and OpenAI tended to talk up their most powerful and capable AI systems. Such systems can cost $100 million or more to train and need a lot of computing power to respond to customer queries or requests.
But many customers balk at the computational costs of using the best models. So in response, AI companies have had to think small.
It is possible to use the different models in tandem, Krieger said. Some companies rely on the most advanced models to set strategy or make plans, but outsource the grunt work to smaller models that can search the Web and synthesize information.
