Downloading takes a few minutes. Operating the model commits a business for months. Behind the promise of open AI models — greater freedom, less dependency, a smaller bill — lies a less spectacular reality: someone has to pay for the servers, monitor the responses and step in when the service goes down. Savings on access to the model do not guarantee savings on running it. Looking ahead to September 2026, trends already observed suggest this distinction will carry more weight in IT departments’ decisions.
Open means neither free nor unconditional
The release of model families such as Meta’s Llama and models published by Mistral AI has changed the available options. A business can download weights, adapt a model and choose its hosting provider, rather than relying exclusively on a remote interface billed by usage. This shift, well underway in 2023 and 2024, made an alternative to proprietary services credible for many projects.
But the word “open” covers several different realities. Accessible weights do not necessarily provide access to training data, every aspect of how a model was built, or unrestricted freedom of use. Licenses vary across families and versions. Some broadly permit commercial use; others impose additional conditions. Before running technical benchmarks, businesses therefore need to read the license, including its redistribution rules and any restrictions.
This check may seem secondary after a successful demonstration. It becomes central when the model is incorporated into a product sold to customers. The legal team must then review the applicable obligations, available warranties and contractual liabilities. The absence of royalties eliminates neither this work nor the risk it seeks to manage.
The server does not disappear with the license
Consider a common case: an assistant tasked with finding information in internal documentation. At the prototype stage, a few users ask questions and receive satisfactory answers. In production, several teams connect simultaneously, documents grow longer and everyone expects a quick response. The challenge is no longer simply to run the model, but to maintain service quality.
Requirements depend on its size, the memory needed, the length of exchanges and the number of concurrent requests. Reducing the numerical precision of the weights through quantization can lighten infrastructure requirements. Batching requests can improve efficiency. These optimizations, however, involve trade-offs: quality must be checked, latency monitored and additional engineering work undertaken.
The decisive factor is often resource utilization. Reserved capacity that sees little use remains costly. Conversely, a stable, heavily used and properly optimized service can make self-hosting attractive. There is therefore no universal threshold beyond which an open model becomes cheaper: it all depends on traffic, availability commitments and the hardware actually deployed.
The hidden bill often has a first name
A superficial comparison pits the price of an API against the monthly cost of a few accelerators. What it leaves out is people. Deploying the inference engine, patching its dependencies, monitoring performance, organizing backups and handling incidents all take time. Even where a platform team already exists, these tasks consume capacity that cannot be devoted to something else.
Maintaining the model itself adds another layer. A new version may provide better answers but change the output format or fail on cases it previously handled well. Evaluations must be rerun, regressions checked and a rollback planned. Adapting the model to business data is no more a one-time operation: products, procedures and document repositories evolve.
A business that already has infrastructure and an experienced team starts with an advantage. For a small organization, buying that expertise can swallow the anticipated savings. Between the two, managed hosting offers a compromise: the customer still chooses the model, while a provider handles part of its operation. The expense moves to a different budget line; it does not evaporate.
Security: keeping the data is not enough
Hosting a model within an organization’s own environment can make data flows easier to control. That is a compelling argument for certain sensitive documents. But location is only one dimension of security. Access rights, technical secrets, conversation logs, updates and the provenance of downloaded components must also be managed.
An assistant connected to a document repository can disclose information to the wrong person if permissions are not properly carried through. Malicious content can attempt to override its instructions: this is the problem of prompt injection. If the assistant has tools capable of sending messages or modifying records, the consequences go beyond an inaccurate answer.
These risks also exist with a proprietary API. The difference lies in particular in how tasks and responsibilities are divided. Self-hosting transfers more operational obligations to the business; it is not a security certification. Isolation, adversarial testing, monitoring and human validation must be included in the budget according to how critical the service is.
Compare a useful outcome, not a listed price
Total cost of ownership must encompass preparation, operation and any eventual exit. Comparing only the price per million tokens obscures differences in quality. A cheaper model that requires three attempts or more human corrections may cost more per completed task. Conversely, a small specialized model may suffice where a large general-purpose model would be excessive.
A rigorous economic assessment must therefore measure several factors using the same tasks:
- Practical quality: acceptable answers, critical errors and time spent on human verification.
- Service delivered: response time, availability and behavior during traffic peaks.
- Full expenses: compute, storage, networking, integration, security, support and staff time.
- Ease of switching: the effort required to change versions, hosting providers or vendors.
The exercise benefits from distinguishing between several traffic scenarios. The average cost of a quiet day says nothing about the capacity required during a commercial launch. Supporting components must also be accounted for: document retrieval, indexing, filtering and observability. The model is often just one component of the service being sold or used.
Freedom still has economic value
Reducing the debate to a battle over prices would nevertheless be a mistake. The ability to modify a model, run it offline or negotiate with several hosting providers has value. This autonomy can justify additional costs, particularly when reliance on a single vendor poses a strategic risk. But it must be distinguished from immediate savings, and businesses must verify that they can exercise it in practice.
What next? For September 2026 and beyond, one plausible path is toward hybrid portfolios: open models for well-understood tasks, managed services to absorb peaks or access certain capabilities. This development is not guaranteed. The best-prepared businesses will be those that measure the cost of a reliable outcome, test their ability to exit and treat operations as a long-term product commitment — not as free after-sales support for a download.