Santander says smaller AI models cut one banking task from about €5 to €0.03. What can we check?

On October 19, 2026, Banco Santander and NVIDIA announced a new stage of their AI collaboration, centered on the Santander AI Lab. It is a useful counterweight to the idea that bank AI means one huge chatbot. The pitch is the opposite: match each task to a smaller model and serve it cheaply.

What Santander says it built

The release describes three pieces. Model Arena is an internal setup for evaluating open-weight models against banking tasks, to find which model suits which workload. Specialized models are tuned with NVIDIA's NeMo Customizer. And the bank is running inference inside its own architecture, using NVIDIA Dynamo-Triton, which it credits for the lower serving cost.

Santander also says it is using NVIDIA Kumo, a family of open-weight models for structured data, to turn tabular bank data into predictions for agent-based services. The Lab also lists adversarial testing of conversational AI in controlled environments.

The two numbers

Santander states that in selected experimental use cases, moving to smaller, specialized models reduced serving costs from around €5 per interaction to approximately €0.03, while maintaining strong performance. Separately, in an experimental customer-service use case it says it improved conversation quality while reducing costs by more than 50%.³

These two figures sit side by side in the release but are not the same measurement. One is a cost per interaction in selected cases. The other is a percentage cut in a customer-service test. Neither comes with the number of conversations, the quality metric, the comparison model or the period.

What we can and cannot check

We can check that the claims were published on Santander's own press page and attributed to its AI Lab. We cannot check how "strong performance" was scored, which general-purpose model produced the €5 baseline, or whether the €5 included costs other than the model's own serving. The release says "selected" and "experimental" in both cases, which means these are not bank-wide production results.

A €5 baseline is also unusually high for a single model call, so it likely reflects a heavy workflow rather than one chat message. The release does not say. We would not compare it with per-token prices from model vendors.

The bigger signal

The more durable statement is strategic. NVIDIA's Aser Blanco says cost per intelligence output is a key metric and that the issue is "not about model size, but about training and fine tuning." That is a vendor speaking, and it sells the hardware. Santander's chief AI officer Iñaki Bernal frames the goal as turning advances into practical banking capabilities.

The three-year framework covers foundation models, AI agents, fraud and anti-money laundering capabilities, real-time decisioning and simulation, including a banking digital twin. It is a commitment to work together, not a result.

Why the cost number is plausible, and why it is not proof

The direction is credible. A narrow task such as classifying a customer message or summarizing a record does not need the most capable model, and serving a smaller model costs less per call. Santander describes moving progressively from large general-purpose models to smaller specialized ones, which is the standard route to lower cost.

What the release cannot tell us is the size of the quality trade-off. "Maintaining strong performance" could mean parity with the large model on a narrow test set, or a small drop that the bank judged acceptable. Banks also carry costs the serving figure leaves out, such as fine-tuning, evaluation, monitoring and the staff who run them.

A useful reading rule: a cost per interaction tells you about the price of one call, while a customer-service saving tells you about the process. The first can be true while the second is small, and the release gives only the first at full precision.

What would make it checkable

Publication of the evaluation design would do it: the tasks, the scoring method, the baseline model and the sample size. Banks rarely publish this, so the useful habit is to treat cost-saving claims in announcements as hypotheses. They tell you where the industry is heading. They do not yet tell you how much money AI has saved.