As companies expand their use of artificial intelligence, the economics of choosing between hosted proprietary models and open-weight alternatives are becoming harder to ignore. But the apparent savings from low token prices cannot be treated as proof of lower total operating cost.
Reuters reported on enterprise interest in open-weight models, drawing on Capgemini research. Capgemini also publishes broader work on AI adoption and business perspectives. The publicly retrieved research material does not independently verify all the specific survey counts and cost percentages described in the Reuters story. We therefore do not repeat those figures as established facts here.
What open weights change
An open-weight model makes trained parameters available for organisations to operate under an applicable licence. This may give a buyer more control over hosting arrangements and software customisation than a purely closed API. It does not mean every model is unrestricted open-source software or that self-hosting is always less expensive.
A proprietary API may reduce operational work, particularly for occasional requests. Self-hosted inference may be attractive for high, predictable volume but introduces hardware, utilisation, maintenance, security and engineering costs. The choice also depends on model quality for the specific job.
A useful cost comparison
Businesses should measure cost per correct, completed task rather than price per million tokens alone. A lower-cost model that repeatedly fails and requires human correction may cost more than a higher-priced model with better reliability. The same principle applies to agents that make multiple tool calls to complete a request.
Privacy rules, regional data requirements, licensing terms and integration complexity can change the decision. A realistic pilot therefore tests representative requests, failures, staff time and infrastructure costs over a defined period.
What the evidence cannot tell us yet
Survey responses about planned adoption are not deployment records. Reported list prices are not actual enterprise contracts. Aggregate public traffic shares are not equivalent to realised revenue or customer savings. The specific percentages reported in secondary coverage require an accessible original study before they can support precise conclusions.
Our assessment: More organisations may find a mixed model portfolio useful, selecting different systems for different workloads. That is a testable purchasing strategy, not a universal cost-saving result.
Original AI-generated conceptual illustration. Credit: Nextwith.ai / OpenAI image generation.