IBM CEO says open-weight models now hit about 90% of frontier performance

On October 27, 2026, Solidot reported remarks by IBM CEO Arvind Krishna on the state of open-weight models. Krishna described how the picture has changed: back in 2024, closed-source frontier models were clearly ahead, scoring roughly 70 on benchmark tests while open-weight models managed only 20 to 30; by late 2025, the gap had narrowed to 70 versus 50. He went on to say that open-weight models today cost only 10% to 25% as much per token as frontier models, while delivering around 90% of their effectiveness. On that basis he argued that for any enterprise the choice is straightforward: given an option that costs 10% of the leading offering while performing at 90% of its level, most companies would avoid paying the extra cost. He therefore suggested that most routine tasks can be handled by open-weight models, with only a minority of tasks handed to frontier models, and that deploying open-weight models inside enterprises carries considerable value. No benchmark data or further documentation from IBM has been made public to back these claims, and no additional reporting has appeared to date.