US bans DeepSeek at civilian agencies as new version launches

China's DeepSeek AI model is facing headwinds in the US, but development continues at the same time. Here are the two stories.

Ban at civilian agencies

The US Department of Homeland Security (DHS) has ordered DeepSeek products removed from civilian federal agencies' IT systems. This also applies to third-party services that use the DeepSeek model. Civilian agencies are also barred from acquiring such products or services, Reuters reports.

DHS points to several threats. According to the department, there is a risk that sensitive data will be transferred to China and that intelligence could be gathered on American individuals and agencies. There is also concern that the models themselves could contain malicious software (malware) or other security holes.

The order covers civilian agencies, not the Defense Department. The Commerce Department and the Navy had previously introduced their own bans. A memo from July asked federal agencies to monitor traffic related to DeepSeek risks. It was made public at the same time as the ban. Reuters gives no deadline for the removal.

New version of the Flash model

On October 17, DeepSeek released an update called V4.1-Flash-1017. According to DeepSeek, it is better at reasoning, coding, agent tasks (where the AI carries out several steps on its own) and image understanding, and it solves tasks faster and more cheaply. It is live on DeepSeek's website, app and API.

All figures below come from DeepSeek itself and have not been independently verified. According to them, the model is ahead of Claude Opus 5 on Agents' Last Exam (35.4 vs. 28.6) and Automation-Bench V1.0.6 (60.5 vs. 50.3). Opus 5, for its part, leads on Humanity's Last Exam (56.3 vs. 42.6/44.0) and on the multimodal tests by 1-2 points.

On price, the difference is large, according to DeepSeek's own analysis. On the DeepSWE v1.1 test, the model scores roughly 77% at under USD 0.25 per task at the highest reasoning level. Claude Opus 5 requires over USD 10 per task for comparable performance.

Background

The original V4.1-Flash was released on September 10 with a 1 million token context window, meaning room for a lot of text at once, and weights under an MIT license on Hugging Face. From September 14, requests to the older V4-Pro are forwarded to V4.1-Flash at Flash prices, until a V4.1-Pro possibly arrives. No date has been announced.

The sources do not make clear whether the ban also affects companies that supply federal agencies, beyond the third-party services built on the model.