DHS ordered DeepSeek purged from federal civilian systems, including third-party tools built on it. Here's what the order covers and who it affects.
What is the DeepSeek federal ban and who does it apply to? #
The Department of Homeland Security issued an exclusion order directing US federal civilian agencies to remove DeepSeek from government computers and networks, and the order extends beyond the DeepSeek app or API itself to cover third-party products built on top of DeepSeek’s models. In practical terms, that means a vendor selling software to the federal government can’t quietly route requests through DeepSeek on the backend. The order applies to federal civilian systems, not to private companies, individual developers, or everyday consumers, who remain free to download, run, or build on DeepSeek’s models.
TL;DR #
- DHS ordered DeepSeek removed from US federal civilian computers and networks, and the scope reportedly reaches third-party software built on DeepSeek models, not just the official app.
- The ban is a government procurement and security policy, not a criminal law, so it is not illegal for ordinary people or private businesses to use DeepSeek in the US.
- DeepSeek’s appeal is almost entirely about cost and openness: its models are released as open weights under the permissive MIT license, letting anyone download, modify, and run them without paying DeepSeek anything.
- The newest model discussed around this news, DeepSeek V4.1 Flash 1017, posted benchmark results that beat Claude Opus 5 and GPT 5.6 Sol on several coding and agent tests while costing a small fraction as much per task.
- Independent testing from groups like Artificial Analysis and Arena found DeepSeek sits in the most cost-efficient zone on their charts, even though it trails the top closed models on raw intelligence.
- The same independent testing flagged a real weakness: a high hallucination rate on factual questions, even as the model performed well and reliably on agent and coding tasks.
- The security concern driving the federal order is less about the model’s answers and more about data handling, infrastructure control, and the fact that DeepSeek is a Chinese company subject to Chinese law.
Remy doesn't build the plumbing. It inherits it. #
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Why did DHS move against DeepSeek specifically? #
The federal government’s objection to DeepSeek isn’t about whether the model writes good code. It’s about where the data goes and who controls the infrastructure behind it. DeepSeek is a Chinese company, and US officials have raised concerns common to this entire category of foreign-built AI tools: that prompts, documents, and usage patterns sent to a Chinese-operated service could be accessible to the Chinese government, and that embedding such a model inside federal workflows creates a dependency on infrastructure the US doesn’t control.
This mirrors earlier federal actions against other foreign-linked software and hardware, where the concern wasn’t a single confirmed breach but the structural risk of routing sensitive government activity through systems outside US jurisdiction. The order treating third-party products as covered, not just DeepSeek’s own app, closes an obvious loophole: an agency can’t comply by avoiding the DeepSeek brand name while still running DeepSeek’s models under the hood through some other vendor’s product.
The timing adds urgency to the policy conversation. Security researchers have also reported incidents of bad actors using DeepSeek’s models as part of tooling in cyberattacks, which reinforces the government’s argument that the risk isn’t hypothetical. None of this says anything bad about the model’s coding ability. It’s a separate axis entirely: capability versus trust in the operator.
Is it illegal for regular people to use DeepSeek? #
No. The DHS order is a federal procurement and cybersecurity policy that governs what agencies can run on government systems. It does not create a criminal prohibition on using DeepSeek, and it does not apply to private citizens, academic researchers, startups, or enterprises outside the federal government. Because DeepSeek publishes its models as open weights on Hugging Face under the MIT license, anyone can legally download them and run them on their own hardware, anywhere, with no DeepSeek account, no data sent to any server, and no licensing fee.
That openness is actually central to why the policy question is awkward. A closed, hosted ban is easy to enforce: block the website, block the API endpoint. An open-weight model is just a file. Once it’s downloaded, it can be run entirely offline, modified, fine-tuned, and redistributed, and no single company or government can revoke access to it. The federal order can stop agencies from calling DeepSeek’s hosted API or buying software that quietly depends on it, but it can’t stop a private company from down the weights and running the model on its own servers with no connection to DeepSeek at all.
What makes DeepSeek’s models competitive on performance and price? #
DeepSeek’s pitch has consistently been value, not supremacy. The company’s recent release, V4.1 Flash 1017 (an update to the V4.1 Flash model, with “1017” marking the October 17 release date), is the smallest model in its V4.1 family, built to be fast and cheap, and it still produced benchmark scores that landed near or above some of last generation’s flagship American models.
Remy doesn't write the code. It manages the agents who do. #
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
On DeepSeek’s own reported figures, the model scored higher than Claude Opus 5 and GPT 5.6 Sol on several coding and agent benchmarks, including a software-engineering test (77% vs. 74% and 73%), a business-workflow automation benchmark, and an “Agent’s Last Exam” test where it led every model in the comparison. It trailed on others, including Humanity’s Last Exam and a few of the hardest coding and vision benchmarks, where Opus 5 stayed ahead.
The pricing gap is where the story gets dramatic. DeepSeek’s API pricing has been reported around 30 cents per million input tokens and $1.20 per million output tokens, compared with roughly $5 and $25 for Claude Opus 5, and around $4 and $20 for GPT 5.6 Sol, a difference in the range of 15 to 20 times. Independent evaluators found similar patterns: Artificial Analysis placed the model in what it calls the most cost-attractive quadrant of its intelligence-versus-price chart, and Arena ranked it the top open model in its Agent Arena leaderboard, ahead of every other open model and behind only closed systems from Anthropic, OpenAI, and Google.
It’s worth noting that DeepSeek’s own comparisons leaned on Claude Opus 5 and GPT 5.6 Sol, which were not the newest flagship models at the time of release (Anthropic’s Opus 5.5 and OpenAI’s newer Astra and Soul variants had already shipped). Matching a prior generation’s flagship at a fraction of the cost is still meaningful, but it’s not the same as leading the absolute frontier.
Is DeepSeek actually a security risk, or just a geopolitical one? #
Both, depending on what you mean by risk. On pure output quality, independent testing found a real technical weakness: a high hallucination rate on factual, “I don’t know”-style questions, far higher than competing models like Gemini. That’s a genuine reliability concern for anyone using the model for research or fact-based work, regardless of who built it.
The federal ban, though, is driven by a different kind of risk: data sovereignty and operational control. Even a technically excellent model raises concerns if queries, documents, or agent actions pass through infrastructure operated by a company subject to a foreign government’s legal authority. That risk exists independent of hallucination rates or benchmark scores, and it’s the reason the order targets usage and procurement rather than the model’s technical merits.
Interestingly, in the specific area the government might worry about most, agentic tool use, independent testers reported no issues with “tool hallucination” (the model inventing commands or tools that don’t exist), suggesting it behaves more conservatively in structured agent tasks than in open-ended factual chat.
Frequently Asked Questions #
Can US companies still legally use DeepSeek?
Yes. The DHS order governs federal civilian agencies and their vendors, not private businesses. Companies outside the federal government can use DeepSeek’s hosted API or download its open-weight models without breaking any law.
Does the ban cover apps built using DeepSeek’s technology?
Reports indicate the order’s scope extends to third-party products built on DeepSeek models, not just DeepSeek’s own branded app, meaning agencies need to check what’s running underneath the software they procure, not just the vendor’s name on the box.
Why is DeepSeek so much cheaper than American models?
DeepSeek uses a mixture-of-experts architecture, where only part of the model activates for any given request, which lowers compute costs. Combined with aggressive pricing strategy and free open-weight distribution, this lets DeepSeek undercut closed American models by a wide margin on a per-token basis.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Does open weights mean DeepSeek has no privacy risk?
Down and running DeepSeek’s weights locally means no data leaves your own hardware, removing the data-transmission concern entirely. The privacy and security risk applies specifically to using DeepSeek’s hosted service, where prompts are sent to DeepSeek’s own servers.
Is DeepSeek actually better than Claude or GPT models?
It depends on the task. DeepSeek’s recent model beat older versions of Claude Opus and GPT on several coding and agent benchmarks while costing a small fraction as much, but it trailed on the hardest reasoning tests and had a notably higher hallucination rate on factual questions.