LinkedIn swapped out some OpenAI models to keep its AI costs flat

The Information reported on Monday that LinkedIn has flattened the cost of its AI Hiring Assistant. It did this by swapping out some OpenAI models. The outlet says the swap held the line even as use of the tool rose over the past year. The story is by Aaron Holmes, and the rest is behind The Information's paywall.

LinkedIn is owned by Microsoft. Hiring Assistant uses large language models to automate candidate search and outreach for recruiters. The Information frames the move as part of a wider pattern. As more people use AI features, the companies that sell them spend more on servers. Margins get tighter. Some firms are finding ways to hold the cost line.

What the swap signals

The detail of which models got swapped, and where, sits behind the paywall. The point in the visible summary is the shape of the choice. A feature whose use is rising. A cost that stays flat. A model change as the reason.

The move matters in part because of who is doing it. OpenAI is Microsoft's close partner on AI. LinkedIn is Microsoft-owned. A product this visible can save money by taking some work off an OpenAI model. Other teams can ask the same question. Which step in a pipeline needs the strongest model? Which step can run on a smaller one, or a cheaper vendor? The answer that worked last quarter may not be the one that works today.

LinkedIn's own help pages have said for months that Hiring Assistant uses OpenAI models through Microsoft Azure. It also uses LinkedIn's own models. The product is one of Microsoft's most visible AI-in-the-workflow releases. Recruiters use it to screen candidate pools and draft outreach at scale. A tool like this already costs real money. A rising use curve is where a model mix starts to matter.

What a team should check this week

A team that runs AI features on thin margin can do a cost-and-model audit. List every call the product makes. Mark the model each call uses and the token count it averages. Ask which of those calls still need the most expensive model. Candidate routing, outreach drafts, long-document steps and chat replies rarely all need the same model. The LinkedIn story suggests a real-world case where moving some of them cut the bill without cutting the use.

A good audit also keeps a quality bar for each step. The cheaper model is only cheaper if the user does not notice. One way is a small eval set per call. Run both models on the same prompts. Judge the answers side by side. Keep the swap only where the cheaper model matches on the cases that matter.

The Information does not name the LinkedIn calls that changed, or the new models. It does not say the share that went to cheaper OpenAI models, or to Azure or LinkedIn's own models. Those answers may come later, from a Microsoft engineering post. Until then, the useful part is the shape of the move. A swap on this scale is now worth doing.

Source

The Information by Aaron Holmes. The article sits behind a paywall.