AI TRENDS | LinkedIn Replaces Some OpenAI Models, Keeps Server Costs Flat

LinkedIn said operating costs for its AI recruiting assistant remained stable over the past year despite rising usage. According to Sina Finance, more than 25,000 companies were paying to use the tool as of late last month, up from 20,000 in data disclosed in September. LinkedIn CTO Eran Berger said the company expects the tool’s user base to grow but does not plan to increase GPU spending during the fiscal year ending in June 2027. The assistant helps recruiters screen candidates and contact them. LinkedIn has shifted some tasks from OpenAI’s general-purpose models to smaller, custom models developed through knowledge distillation, in which smaller models are trained to mimic the output quality of larger models for specific tasks. Engineering Vice President Prashanthi Padmanabhan said that a year ago, most of the assistant’s features relied on OpenAI models. Candidate searches, ranking candidates against job requirements, and drafting outreach messages now run on smaller models. She said their output is close to that of OpenAI models and can reduce operating costs by up to 67% compared with using OpenAI models directly a year ago. OpenAI has since cut prices for several models and released more efficient versions; LinkedIn did not say whether its models remain less costly than OpenAI’s current offerings. Padmanabhan said new AI features still initially use OpenAI models, which LinkedIn later replaces with custom smaller models. LinkedIn and other companies within Microsoft do not pay model licensing fees for OpenAI models under Microsoft’s commercial agreement with OpenAI, but do pay computing costs. Other companies paying the full cost of OpenAI or Anthropic APIs could achieve greater savings by switching to third-party smaller models. The agreement also allows LinkedIn to distill OpenAI models, while OpenAI prohibits other companies from doing so to build competing products. AT&T and Cursor have also reported success with customized open-source models without using distillation.