Banco Santander said on October 19, 2026 that smaller, specialised AI models cut its serving costs from around €5 to about €0.03 per interaction in selected experiments, and it signed a three-year collaboration framework with NVIDIA to expand the work. Why it matters: a global bank is publishing hard per-interaction cost numbers for AI — and they favour small, task-specific models over large general-purpose ones.
Madrid and Santa Clara — October 19, 2026.
What happened?
Santander announced a new milestone in its artificial intelligence collaboration with NVIDIA. Through its Santander AI Lab, the bank has built new capabilities to evaluate AI models and agents for banking applications, work out which model fits each task best, and deploy that model in the most efficient way. Alongside the results, the two companies established a three-year framework covering the next phase of the collaboration.
What did Santander actually build?
The Lab has been moving the bank away from large general-purpose models towards smaller, specialised models designed for specific banking tasks. It built Model Arena to evaluate open-weight models against banking tasks and identify the right model for each workload, and it is also developing Model Arena to design, simulate, evaluate and supervise banking agents. Models are being specialised with NVIDIA NeMo Customizer, and the Lab has built specialised intelligence for core banking workloads on NVIDIA Kumo, a family of open-weight foundation models for structured data. Inference is being industrialised inside Santander’s own architecture using NVIDIA Dynamo-Triton, which the bank credits for the serving-cost reduction. The Lab also does performance engineering and adversarial testing of conversational AI in controlled environments before capabilities scale across the bank.
What were the results?
In selected experimental use cases, Santander says the specialised-model approach reduced serving costs from around €5 per interaction to approximately €0.03 while maintaining strong performance. In one experimental customer-service use case, it says it improved conversation quality while reducing costs by more than 50% by deploying a more efficient model. These are the bank’s own experimental figures, not audited production results, and the release does not say how widely the approach is deployed yet.
What does the three-year framework cover?
The next phase covers foundation models, AI agents, fraud and anti-money laundering capabilities, real-time decisioning and advanced simulation technologies, including a banking digital twin. Santander describes the direction as a move from open-weight models towards specialised models and governed banking agents built for greater control, safety and lower operating costs at scale.
Why it matters
Most enterprise AI announcements offer no numbers at all. Santander’s do, and they point the same way recent enterprise cost stories have pointed: the expensive part of AI at scale is serving it, and the fix is often a smaller model matched to one job, tested against real tasks, rather than the biggest model available for everything. For readers weighing AI tools for their own businesses, that is the practical takeaway — measure cost per finished interaction, not cost per token in isolation. Everything AI, tested: we will treat the €0.03 figure as a vendor-reported experiment result until Santander reports it in production.
Frequently asked questions
What did Santander announce on October 19, 2026?
A new phase of its AI collaboration with NVIDIA: evaluation and specialised-model capabilities from its AI Lab, plus a three-year framework covering foundation models, agents, fraud and anti-money laundering, real-time decisioning and simulation, including a banking digital twin.
How much did Santander say specialised models cut its AI costs?
From around €5 to approximately €0.03 per interaction in selected experimental use cases, and by more than 50% in one experimental customer-service case, where conversation quality also improved.
What is Santander’s Model Arena?
A system built by the Santander AI Lab to evaluate open-weight models against banking tasks and match the right model to each workload; the bank is also developing it to design, simulate, evaluate and supervise banking agents.
Sources: Santander press release (October 19, 2026), PYMNTS.