IBM CEO Puts a Number on the Open-Weight AI Value Proposition: 90% of the Performance at 10% of the Token Cost

The question of whether enterprises should pay a premium for the most capable AI models now has a concrete data point from one of the industry’s largest vendors. IBM CEO Arvind Krishna told financial analysts during the company’s latest earnings call that open-weight models have reached roughly 90% of the effectiveness of frontier closed models—while costing between 10% and 25% per token.

The claim, first reported by Xataka, came during the Q&A session of IBM’s earnings presentation. RBC Capital Markets analyst Matt Swanson asked how large companies are reconciling two pressures: controlling AI spending and maintaining control over deployments. Krishna’s response traced the evolution of open-weight alternatives over the past two years, framing a cost-performance shift that he argues is reshaping enterprise procurement decisions.

Key facts

Krishna’s timeline of the gap

Krishna described a narrowing gap that he said reached an inflection point. At the end of 2024, he told analysts, a frontier model might score 70 points on a broad evaluation, while an open-weight alternative scored 20 to 30. “It’s too big a gap. Nobody is going to accept it,” he said. By the end of 2025, the frontier still scored around 70, but open-weight models had climbed to about 50. Now, he estimates open-weight models reach roughly 90% of frontier effectiveness, at a fraction of the per-token cost.

Krishna did not name specific benchmarks or models during the call. The numbers are his own characterizations of the market landscape, not a published benchmark result. IBM develops the Granite family of open-weight models and also maintains a commercial alliance with OpenAI, giving the company interests on both sides of the closed-versus-open debate.

What the numbers mean for enterprise buyers

The cost calculation hinges on how AI services are billed. Providers charge per token—a unit representing fragments of input or output text. A system that costs 10% as much per token can deliver dramatic savings at scale, but only if the quality difference is acceptable for the task.

Krishna argued that many common enterprise use cases do not require a frontier model. He listed examples including summarizing an email, drafting a reply, reviewing a contract, or handling a customer call. “All of that can be done just as well with the much cheaper alternative,” he said. The statement is his assessment of typical workloads, not a guarantee of equivalent results across all scenarios.

The infrastructure trade-off

Running open-weight models on private infrastructure introduces its own cost structure. Krishna estimated that a large enterprise could start with two or three server racks to run these alternatives. That equipment requires capital investment, electricity, cooling, maintenance, and specialized staff. For lower-volume workloads, contracting an API from a closed-model provider may still be the cheaper option, even at a higher per-token rate.

Outside validation and caveats

Krishna’s claims are not unsupported. In May 2026, the OECD published a report that placed open-weight models at roughly 90% of the quality index of closed alternatives at market launch, with a cloud-service price around 20% of the closed equivalent. The report was based on earlier data and carries its own methodological notes.

Stanford’s AI Index 2026 offers a slightly different picture. As of March 2026, the best closed model outperformed the best open-weight model by 3.3% in the Chatbot Arena leaderboard, a ranking based on user preference evaluations. That is a narrower gap than the raw 10 percentage-point difference implied by Krishna’s 90% framing, but it measures user preference rather than benchmark accuracy.

Neither metric establishes a universal capability equivalence. The choice between model types depends on task complexity, reliability requirements, and total cost of ownership at a given usage volume.

The bottom line for AI procurement

Krishna’s argument is not neutral—IBM sells both Granite models and infrastructure for running them—but the underlying trend is real. Open-weight models have improved faster than many observers expected, and the cost gap with proprietary frontier models remains wide. For enterprises, the question is shifting from “which model is best?” to “which model is good enough for this specific task, and at what total cost?”

The answer will vary by organization, workload, and risk tolerance. But Krishna’s 90/10 framing gives procurement teams a number to test against their own use cases.

Source: Xataka – “IBM tiene unas cifras que desafían a los gigantes de la IA cerrada: 90% del rendimiento por solo el 10% del coste por token” (https://www.xataka.com/robotica-e-ia/ibm-tiene-unas-cifras-que-desafian-a-gigantes-ia-cerrada-90-rendimiento-solo-10-coste-token)