Santander announced in a press release dated October 19, 2026, issued in Madrid and Santa Clara, that its AI Lab, working with NVIDIA, has built new capabilities to evaluate AI models and agents for banking. The standout figure comes from selected experiments: after switching to smaller, specialised models, the serving cost fell from around €5 per interaction to approximately €0.03, while performance stayed strong.
Model Arena tackles model selection first
Instead of buying a bigger model first, the bank built an evaluation arena. Its AI Lab created Model Arena to test open-weight models against banking tasks and identify which model fits each workload, then specialises the chosen models with NVIDIA NeMo Customizer. For structured data, the Lab built specialised intelligence on NVIDIA Kumo, a family of open-weight foundation models, turning years of structured banking data into predictive insights for agentic services. On the serving side, it is industrialising inference inside its own architecture with NVIDIA Dynamo-Triton. The cost drop comes from this whole combination, not from any single step.
A customer-service experiment provides checkable numbers
In one customer-service experiment, deploying a more efficient model improved conversation quality while cutting costs by more than 50%. Note that the €5-to-€0.03 figure is explicitly limited to selected experimental use cases in the announcement, not an average across the whole bank. It shows how much selection and specialisation can save on a specific task, not that every banking task can be cut by the same ratio.
The three-year framework reaches beyond hardware
Santander and NVIDIA also set a three-year framework for the next phase, covering foundation models, AI agents, fraud and anti-money-laundering capabilities, real-time decisioning and advanced simulation such as a banking digital twin. Model Arena is also being developed to design, simulate, evaluate and supervise banking agents, alongside adversarial testing of conversational AI in controlled environments, so emerging technology is stress-tested against banking requirements before anything scales across the bank.
The real story is a change in the metric: Aser Blanco, NVIDIA's global banking lead, calls cost per intelligence output a key metric. Once banks select models by that measure, large general-purpose models no longer win by default, and smaller models that can be evaluated on real tasks, specialised, and served cheaply will take a larger share. For other financial institutions, the replicable part is not any single model, but the order of operations: build the evaluation arena first, then decide deployment.
