Up Close
DeepSeek turns a model release into something firms can run and meter themselves
DeepSeek published the weights of its V4.1-Flash-1017 model under an MIT licence for self-hosted use and opened its Search and Fetch APIs to all developers, with billing from 31 October at $0.75 per 1,000 calls for each endpoint.
This appears to matter less as another model update than as a packaging move: one company can now take the same family in two directions at once - run it itself or buy a metered service around it. That is likely to strengthen the position of open-weight models in business use, because the commercial offer is no longer only raw download or only somebody else's cloud API. It also makes Microsoft’s position easier to read: if model access keeps arriving with both self-hosting and priced service layers attached, cloud platforms need to compete on control, governance and deployment rather than on keeping a model exclusive.
IMDEA Networks Institute published a study finding that six of nine examined major AI chatbots send conversation data from web or app interfaces to advertising services, including services such as ChatGPT, Claude and Gemini. Meta revised its commercial terms by removing language that had suggested it could use content from business account domains and URLs for AI training, and said the crawling provisions apply only to its ads and business products.
Taken together, these two moves suggest that trust in AI is becoming an operating condition, not a promise on a product page. If chatbot interfaces are found to leak conversation data into ad pipes, and a platform owner retreats from wording that implied broader training rights over business content, buyers get a clearer reason to separate model capability from the surfaces through which they access it. For Microsoft and everyone else selling AI into work, that is likely to push more value toward governed enterprise channels and away from casual consumer-style interfaces.