Open-source AI promises more control and better value

A company that runs open-source AI on its own servers keeps its data in-house and decides when the software changes. With heavy, steady use it is better value than paying OpenAI, Anthropic or Google for every request, but the best open-source options still score well below the frontier labs’ best.

Open-source AI comes from makers who publish the trained system itself. A company can download it and run it on hardware it controls.

The frontier labs keep their AI closed. OpenAI’s GPT, Anthropic’s Claude and Google’s Gemini run only on their makers’ servers, so a company sends its requests and data there and pays for each request.¹

Mistral, the French maker of open-source AI, urges companies to run its software themselves. It makes two promises.

Data stays local when the AI runs close to it, and skipping the margin the frontier labs build into their prices gives better value for money.²

Both promises hold, but each comes with a condition.

The first, control, depends on whose hardware it runs on. On the company’s own servers, its data stays in-house, and the AI changes only when the company replaces it. A process built on one version keeps working on that version until the company chooses to switch.¹

The frontier labs update or retire their AI on their own schedule.¹

A cloud provider sets the same kind of schedule for open-source AI it rents out. On Microsoft’s cloud platform, Mistral’s own systems have a fixed 12-month lifecycle and stop answering requests after their retirement date.³

Open-source AI gives a company control only when the company runs it itself.

Conditions apply

The second promise, value, depends on the workload. The software is free to license, but running it costs money. A server rented by the hour costs the same whether it is answering requests or sitting idle.

The frontier labs charge per request. Heavy, steady use makes the dedicated server the cheaper option.

Light use makes it the more expensive one, because the company also pays for the idle hours. A company can tell which case it is in only by measuring its own costs.⁴

That arithmetic worries companies still trialling open-source AI. Half of them name the cost of running it themselves as a barrier to going live, up from 40%.⁵

Quality adds a third condition, because both promises matter only for work the open-source option does well enough. In the latest independent ranking, the best system a company can download scores 32, against 58 for Anthropic’s Claude, which tops the ranking. Most of that ranking now rests on private tests, so the gap is an estimate that only the ranking’s makers can check.⁶

In practice, companies already use both. Measured by the volume of text processed, open-source AI now handles a third of their AI usage, up from under a quarter a year ago, and they expect it to reach two fifths within a year.

Most firms running it in production have moved some tasks away from the frontier labs. They host it mostly in a cloud they manage or on their own premises, and most give cost savings as a reason.

Almost all of them say their use of GPT, Claude, Gemini and the like is still growing too.⁵

So the choice is made task by task. Work on the company’s own data, or work that relies on one fixed version, may be better run in-house, where the company controls it. With heavy, steady use it can be better value there too. Tasks that need the strongest AI available can stay with the frontier labs.