In April 2026, Revolut began running open-weight AI models for its engineers. That same month, open models established a sustained majority of token consumption on OpenRouter. The coincidence suggests that several models were becoming capable enough to handle substantial everyday work at prices that encouraged wider use.

Revolut’s AI operations lead, David Abitbol, describes a deployment serving roughly 2,000 engineers. The company initially adopted GLM-5.1 and subsequently moved to GLM-5.3, running on 56 Nvidia B200 GPUs. He reports spending around one-tenth of its previous Anthropic and OpenAI bill, alongside a 70% improvement in engineering lead time.

These are company-reported results. Lead time measures elapsed time from starting a feature to having it merged and ready for deployment, rather than overall productivity. The comparison also combines changes in models, infrastructure and working practices. Nevertheless, it provides a useful example of how organisations can change the economics of AI consumption.

Fifty-six B200s represents the GPU equivalent of seven eight-GPU Nvidia DGX servers: substantial infrastructure, spread across a large engineering workforce. Together’s published cluster rental rates imply roughly $278,000–$335,000 monthly for that capacity running continuously, or $139–$167 per engineer before additional costs. This is an illustrative calculation, not Revolut’s disclosed bill.

Does the saving suggest Revolut previously used frontier AI inefficiently? Possibly. Some tasks may have received more expensive intelligence than they needed. Better model selection, caching and context management might also have reduced the earlier bill. But spending limits created another inefficiency: engineers conserved tokens instead of using AI wherever it could save working time.

LinkedIn offers a related example, with a different mechanism. According to The Information, it has kept the cost of serving its AI Hiring Assistant roughly flat while customer numbers and usage have grown. More than 25,000 businesses now pay for the product, which helps recruiters search for candidates, assess suitability and prepare outreach.

LinkedIn initially relied primarily on OpenAI’s general-purpose models. It subsequently used those models to teach smaller open-source models to perform particular tasks, through distillation. The company reports similar results on those tasks at savings of up to 67% against what it paid to run OpenAI models a year earlier.

The comparison is unusually revealing because Microsoft’s agreement reportedly lets LinkedIn use OpenAI’s models while paying only their computing costs. The reported savings therefore reflect running smaller models, even without eliminating a commercial API markup. However, the figure covers selected tasks and does not establish an equivalent reduction in total product costs.

LinkedIn still uses frontier models for new features, then develops specialised replacements where practical. That sequence suggests a division of labour: general intelligence helps establish what an application can do; cheaper, narrower intelligence performs work that has become sufficiently understood and repeatable. Customer feedback and reliable evaluations become essential to deciding when replacement is worthwhile.

This connects with the discussion of messy jobs. Engineering involves documentation, coordination and investigation. Recruitment involves relationships, judgment and organisational constraints. Those jobs can remain complicated while individual activities become predictable enough to automate economically. A specialist that performs one activity adequately may deliver substantial value without reproducing the flexibility of a frontier model.

The April parallel supports this interpretation. OpenRouter attributes much of DeepSeek’s subsequent growth to agent workloads following V4’s 24 April launch. Revolut’s internal GLM traffic would not appear in those figures. These are separate examples of affordable capability expanding the range of work worth delegating, rather than evidence of one event causing the whole market to change.

There are important limits. LinkedIn’s savings compare its specialists with year-old OpenAI costs, while OpenAI has since improved efficiency and lowered prices. LinkedIn also enjoys unusual distillation rights. Neither case provides a complete comparison covering development, operations, reliability, human correction and the best current hosted alternatives.

Nor does efficiency necessarily reduce overall demand. LinkedIn says savings help fund new frontier-powered features, explaining why total computing costs remain steady. Cheaper execution can make additional applications viable, while premium models retain work where their greater capability justifies the expense. Whether total spending rises depends on how much new demand those improvements create.

The pressure on frontier suppliers is therefore specific. They may enable valuable applications without retaining premium revenue from every mature workflow. Meanwhile, application companies gain an incentive to own customer relationships, feedback and the knowledge needed to build suitable specialists. Revolut and LinkedIn suggest a market becoming more selective about which intelligence it buys, how it runs that intelligence, and where the resulting savings create the most value.