DeepSeek’s New Model Brings Better Scores and Bigger Risks

DeepSeek has released a faster model at strikingly low prices. DeepSeek-V4.1-Flash-1017 arrived on October 17, 2026, promising better coding and faster completion of complex tasks while giving businesses another low-cost, open-source AI option. The catch is familiar now: U.S. government restrictions make deployment choices part of the product decision.

Third-party evaluation firm Artificial Analysis scored DeepSeek-V4.1-Flash-1017 at 45, up from 39 for its predecessor. DeepSWE v1.1 reached 77.0 in software engineering, compared with 74.2 before, while Terminal-Bench 4.0 rose from 31.2 to 42.3.

The evaluation cost per task also fell from $0.49 to $0.26. Through the API, the model costs $0.30 per 1 million input tokens and $1.20 per million output tokens during peak hours, which run from 01:00–04:00 and 06:00–10:00 UTC Monday–Friday, excluding Chinese public holidays.

Those prices put the model near the cheaper end of the listed market. Muse Spark 1.2 / 1.3 costs $0.30 total per million tokens, GPT-6 Luna costs $0.60, while Claude Sonnet 5.5 and GPT-6 Astra each cost $12.00. Low prices tend to attract developers before the lawyers arrive. Technology has a sense of timing.

Commercial Use Meets a Federal Ban

On October 19, 2026, the U.S. Department of Homeland Security disclosed an order from the summer directing federal civilian agencies to exclude and remove all DeepSeek technology. The order also covers implementations supplied by other companies, so changing vendors does not automatically change the model underneath.

Civilian agencies operate under the DHS order, while defense contractors already face a statute that bars DeepSeek from Pentagon contract work. The DHS order excludes specified national security systems, but suppliers cannot treat that jurisdictional distinction as permission to use DeepSeek for national security work.

DeepSeek remains a viable option for purely commercial work that never touches a government contract, provided businesses keep it walled off from government work. Sending export-controlled technical data to DeepSeek’s China-hosted services can require a license, while downloading the weights or running the model through a U.S.-based cloud provider avoids export licensing issues.

The larger question around U.S. businesses employing any Chinese model remains, as Carl Franzen put it. Senators have written a defense-bill provision that would extend the Pentagon’s ban to ten more Chinese AI companies and their models, and the White House has considered limits on Kimi and Qwen models.

Adoption Is Already Underway

DeepSeek’s reach extends beyond benchmark charts. Many businesses already use older versions such as R1-1776 for purposes including political censorship removal, while Saudi Aramco improved operations after installing DeepSeek at its main data center.

Tencent lists V4.1 Flash among the models powering its WorkBuddy office assistant. DeepSeek also released DeepInfer on October 21, 2026, an open-source inference framework that supports its DeepSeek-V4.1-Flash model on Nvidia’s SM100 Blackwell GPUs and Huawei’s Ascend 950 chips.

American businesses can download and run or adapt DeepSeek, but that flexibility does not erase the compliance boundary. Companies selling AI to the government should start de-risking now by inventorying Chinese models, walling off government work, stopping new dependencies, reading contracts, engineering for portability, and keeping commercial options open.

That checklist matters because the model’s appeal is not theoretical. DeepSeek keeps giving businesses reasons to try its low-cost, high-performing, open-source AI models, but every new deployment also raises the cost of discovering too late where the system is allowed to operate.

Stay connected via Google News

Follow us for the latest AI News