IBM CEO says open models hit 90 percent of frontier

An enterprise paying frontier prices for every email summary has one hard sum to run. How much of its work really needs the best model? IBM CEO Arvind Krishna gave his answer on the company's Q3 earnings call. The Register reports that Krishna told investors open weight models now reach 90 percent of frontier scores. He said they do so at 10 to 25 percent of the token cost.

The gap Krishna named

Krishna said that in 2024, a broad benchmark would score a top closed model around 70. An open model would score around 20 or 30. The Register quotes him on the change since. He said that in late 2025, the gap began to shrink. The 70 to 20 or 30 split became 70 to 50. Krishna then said the open models now cost 25 percent to 10 percent of the frontier token price. He added that is attractive, given they are now 90 percent as effective.

What that means for a buyer

Krishna put the choice as one question for any enterprise. Pay 10 percent of the cost for 90 percent of the quality, or pay 10 times more for the best. He said most buyers will not pay the premium. The list of tasks that do not need frontier models is short and ordinary. Summarising an email. Reading a contract. Answering a customer call. Composing an email. The Register quotes him as saying all of these can be done with the cheaper answer.

The hard part Krishna named

The CEO said many businesses can build useful AI with two or three racks on their own site. The hard part is running it. Krishna told the call his question for the enterprise is a simple one. Where are the kernel engineers? Where are the people who worry about the underlying network? Who can make sure the model is clustered on GPUs close to each other on the network? He had no answer, other than that IBM will produce software to help. IBM's own Spyre mainframe accelerator cards are part of that pitch.

The numbers back up a shift that was already visible. Open weight model releases from Moonshot, DeepSeek and Alibaba scored close to the top closed labs this quarter. They did so at a fraction of the inference cost. The Register notes IBM's infrastructure revenue grew 5 percent in Q3 to $3.7 billion. That was slower than Dell and HPE. Red Hat grew 11 percent, with virtualization deals worth $800 million. A buyer taking Krishna at his word still needs to find the engineers he admits are hard to hire. That gap is where the on-prem AI story stops being about model quality. It starts being about who runs the server.

Source

• Simon Sharwood, The Register: IBM CEO says open weight models are now irresistibly good and cheap on-prem