New Delhi: Anthropic said Tuesday it will let its enterprise customers keep the data they feed into its Claude models on their own servers rather than Anthropic’s, a shift the company is making as governments in the US and Europe tighten rules around powerful AI systems.

The change comes alongside the release of two new models, Fable 5.1 and Mythos 5.1. The data set-up is called Enterprise Frontier Safeguards, under which a business stores what it sends to the model on cloud servers it controls, and any human review of that data is done by the customer rather than Anthropic.Anthropic says this gives the privacy of a no-retention deal, while still allowing it to catch misuse, and that it built the arrangement with more than a hundred customers and with Amazon, Google and Microsoft. The set-up reaches enterprise clients starting this autumn.

The launch is also the first from Anthropic to run under a European Union code of practice that the company signed in July. To meet it, the firm now adds a hidden statistical watermark to text produced by its newer models. It says the mark carries no information about the user, and does not change what the model writes. It is opening a tool that can detect the watermark to regulators, researchers and other approved groups, as the code requires.Politics has shaped the release in other ways. Anthropic had shipped the earlier Fable 5 and Mythos 5 in June, and then pulled both within days to comply with US export controls, before switching them back on in July once Washington lifted those controls. The more capable of the two new models, Mythos 5.1, goes only to organisations the company has vetted, and its biology access programme was set up together with the US government. Fable 5.1, which shares the same engine, is open to public.For the developers who pay to use the models, the news implies a lower bill. Anthropic has cut the price of what it calls cache reads, the amount a customer pays when the model reuses text it has already gone through, instead of reading it again. That cost falls by three-quarters, to 25 cents per million tokens. A token is a small piece of text, about a few characters long, that these models count as they work. The rest of the pricing holds at 10 dollars per million tokens fed in and 50 dollars per million produced. Anthropic says the change cuts the running cost by about a quarter for a normal workload, and by close to half for jobs that lean on tools and long context. Those figures come from its own August usage data.Anthropic also says it has loosened some filters that were stopping harmless questions. Fable 5.1 now blocks about 60 percent fewer such cases in cybersecurity, in part because it is now allowed to look for flaws in software, though it is still barred from writing code that would attack them. Questions on advanced biology, and tasks such as break-in testing, get passed to the company’s older Opus models.The company put weight on a set of research claims, which it presented as a sign of where AI is heading. It says Mythos 5.1 designed proteins that lock onto targets in the body, the first step in making many medicines, and that two outside labs found the designs worked in about half the cases across 12 targets. It adds that Fable 5.1 built a height map covering about a third of Venus from radar data that NASA’s Magellan mission gathered in the early 1990s, and that it will release the map free before upcoming NASA and European missions to the planet. In a third case, it says, Mythos 5.1 rewrote code so that seven biology research models ran up to two and a half times faster on NVIDIA chips.On safety, Anthropic says it tested the models for misuse in chemical, biological and cyber weapons before release. Its finding is that the new model is more capable than the last, but still sits below the point that would force stricter controls under its own risk policy. Outside groups ran tests as well.Fable 5.1 is available now on Anthropic’s platform, and through Amazon, Google and Microsoft. Mythos 5.1 goes to a set of US organisations for now, with the company saying that it is working with the US government to open it up more widely.