Google introduces new class of cheap AI models as cost concerns intensify

San Francisco: Alphabet's Google on Wednesday announced updates to its Gemini family of large language models, including a new product line with competitive pricing to low-cost artificial intelligence models like that of Chinese rival DeepSeek.

The tech giant offers several versions of Gemini that vary in price and performance. It already offered a lightweight variant known as "Flash" but its new "Flash-Lite" model is even cheaper.

On Wednesday, Google released Gemini 2.0 Flash to the general public after previewing it to developers in December. It also launched Flash-Lite and released a new version of its flagship "Pro" model into test phases.

Google created Flash-Lite after receiving positive feedback about the 1.5 version of Flash, Koray Kavukcuoglu, chief technology officer of Google's DeepMind AI lab, said in a press release. Gemini 2.0 Flash is costlier than its predecessor.

The cost to develop AI models and, in turn, the cost to use them have come under investor scrutiny in recent weeks after DeepSeek revealed it spent less than $6 million on the final training run of a model. Developers at leading U.S. AI firms said the total cost was likely magnitudes larger.

Still, DeepSeek's rise drew questions on the earnings calls of Alphabet and rivals Microsoft and Meta. All have so far signaled intent to continue massive capital expenditures in the field.

Alphabet shares slumped Tuesday in part due to investor pessimism around a planned capex hike that was 29% higher than Wall Street expected.

Certain inputs on Gemini Flash-Lite cost $0.019 per 1 million tokens, a term for the units of data processed by an AI model. That compares to $0.075 on the cost-efficient version of OpenAI's flagship model and $0.014 on DeepSeek's cheap model, though DeepSeek states on its website that the pricing will increase fivefold on Feb. 8.