DeepSeek V4.1-Flash-1017 Review: Low Price, US Agency Ban
DeepSeek's newest open-weight model scores 45 on the independent Artificial Analysis index at low API prices, yet US federal civilian agencies must now remove DeepSeek products. Capability is rising while access is narrowing.
Key takeaways
- DeepSeek released V4.1-Flash-1017 on 17 October 2026, with weights on Hugging Face under an MIT licence [1][2].
- Artificial Analysis scores it 45 on its index (v4.3.2), up from 39 for the September V4.1-Flash, but it used 210 million output tokens to run the index against a median of 140 million for comparable models [4][5].
- The coding and agent gains are DeepSeek's own numbers. In DeepSeek's table, Opus 5, the comparison model on its card, still leads on four of six rows [2].
- The Department of Homeland Security (DHS) order, public on Monday 19 October, bars civilian agencies from DeepSeek products. China's foreign ministry objected on 20 October [7][8][11][12].
What is new
DeepSeek says 1017 improves benchmark results, finishes tasks faster and writes shorter answers than the September release. It credits reinforcement learning with self-feedback for building whole software projects, and says it is tuned for DeepSeek Harness, CodeBuddy (Tencent) and WorkBuddy [1]. The model card calls it a minor update, with image and text input, a 1-million-token context window and about 763 billion stored parameters [2].
Size matters for self-hosting. The repository is 510 GB, and the card ships a Python reference but no Jinja chat template, recommending DeepSeek's own Rust library for production [2][3]. DeepSeek also opened its Search and Fetch APIs to all users [1].
Benchmarks: vendor claims versus independent data
All figures below come from DeepSeek's model card, run by DeepSeek, not independently verified [2].
| Benchmark | 1017 | Earlier Flash | Opus 5 |
|---|---|---|---|
| Humanity's Last Exam (no tools) | 42.6 | 36.8 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 74.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 75.3 |
| Apex Bench v1.1 | 72.1 | 64.6 | 65.8 |
| BabyVision (with tools) | 92.5 | 89.6 | 94.1 |
The independent check is Artificial Analysis. It rates 1017 at 45, runs it at about 211 tokens per second with 1.09 seconds to first token, and estimates $0.26 per index task [4]. Its page does not show separate coding or agent scores for this model, so the coding claim still lacks outside confirmation.
Pricing and availability
DeepSeek's API charges $0.15 per million input tokens and $0.60 per million output tokens off-peak, and double that at peak (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday). Cache hits cost $0.003 off-peak [6]. Artificial Analysis lists the peak rates, $0.30 and $1.20 [4]. DeepSeek's larger V4-Pro costs $0.66 and $1.98 off-peak [6].
Search and Fetch cost $0.75 per 1,000 calls each. Calls made through DeepSeek Harness, Claude Code or Codex are free until 31 October, 16:00 UTC [1]. The model runs on DeepSeek's website, app and API, and as downloadable weights on Hugging Face and ModelScope [1].
Strengths and limits
Strengths: very low price, an MIT licence that allows self-hosting, a 1-million-token window, and fast output [2][4]. The DeepSeek release note does not state the licence; the model card does [1][2].
Limits: token appetite offsets some of the price advantage, since the index run used about 50% more output than the median [4]. On harder reasoning (Humanity's Last Exam) DeepSeek's own table shows a 13.7-point gap to Opus 5 [2]. DeepSeek's release note lists no known limitations, and its demos are curated showcases [1].
Best fit: high-volume coding and agent workloads outside US federal work, where cost per task matters. Who should wait: teams that need independent coding results, strict data-residency guarantees or US federal compliance.
The US order and China's response
DHS ordered federal civilian agencies to stop buying and remove DeepSeek products, plus third-party products built using DeepSeek implementations, citing the Federal Acquisition Supply Chain Security Act of 2018 [8][9]. It does not impose a general ban on US businesses or individuals [8]. The notice dates to late July, according to AP, and was made public on 19 October [11]. Gizmodo reports that Homeland Security Secretary Markwayne Mullin's note also warned that the models could contain malware, which is one outlet's account [10].
DHS cites Chinese laws that allow state access to private company infrastructure, data collected at account creation, and risks from user prompts [9]. The Register says it is unclear whether the order covers API access or US-hosted copies of the open weights [7]. On 20 October Chinese foreign ministry spokesperson Mao Ning said China opposes stretching the concept of national security and politicising technology, and will defend Chinese companies' rights, as AP and state broadcaster CGTN report [11][12]. DeepSeek had not responded to AP's request for comment [11].
The global picture
United States: the order extends earlier bans by Commerce, the Navy and several states [7][9]. China: officials object, and Beijing has separately proposed export controls on overseas deployment of some Chinese AI models, covered in our 17 October article [10]. South Korea, Australia, Taiwan and Italy: each restricted DeepSeek on government systems or in other ways earlier [8][9]. Europe, Japan, India, Gulf: no new decision found in our checks.
What to watch next
- Order scope: how agencies read "developed using DeepSeek" for gateways such as OpenRouter [7].
- Independent tests: Artificial Analysis coding scores and third-party agent evaluations of 1017 [4].
- Billing: Search and Fetch charges begin after 31 October, 16:00 UTC [1].
- Diplomacy: AP expects a Trump–Xi meeting in mid-November, with AI on the agenda [11].
Our take
The price and licence make 1017 hard to ignore, but treat the headline scores as claims until outsiders reproduce them. The order is a procurement rule for federal civilian agencies, not a technical verdict; both facts can be true at once. For a fuller model comparison, compare cost per finished task, not per token.
FAQ
Is DeepSeek V4.1-Flash-1017 open source? Its weights are published under an MIT licence on Hugging Face, per the model card [2]. We did not review the training data.
Does the US order ban DeepSeek for everyone? No. It applies to federal civilian agencies and their suppliers, not to private businesses or individuals [8].
How much does it cost? From $0.15 per million input tokens and $0.60 per million output tokens off-peak, double at peak [6].
Sources
- DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real, DeepSeek API docs, 17 October 2026. Tier 1.
- DeepSeek-V4.1-Flash-1017 model card, Hugging Face, October 2026. Tier 1; benchmark figures are DeepSeek's own.
- DeepSeek-V4.1-Flash-1017 files, Hugging Face, read 20 October 2026. Tier 1.
- DeepSeek V4.1 Flash 1017 (max), Artificial Analysis, read 20 October 2026. Independent evaluation; Tier 1 for its own data.
- DeepSeek V4.1 Flash (max), Artificial Analysis, read 20 October 2026. Earlier September version scored 39.
- Models & Pricing, DeepSeek API docs, read 20 October 2026. Tier 1.
- DeepSeek deep-sixed at all US federal civilian agencies, The Register, 20 October 2026.
- US orders federal agencies to remove China's DeepSeek AI products, Korea JoongAng Daily, 20 October 2026.
- DHS bans DeepSeek across federal government networks, PYMNTS, October 2026. The article labels its posting date inconsistently; we use the Monday 19 October date given by The Register, Gizmodo and JoongAng.
- Trump Admin Bans DeepSeek From Federal Agencies, Gizmodo, October 2026. Single-outlet claims labelled in the text.
- China hits back at US ban of DeepSeek from government systems, AP via KIRO 7, 20 October 2026.
- China slams US ban on DeepSeek AI tools, CGTN (Chinese state media; official position only), 20 October 2026.
AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. Research and drafting are assisted by AI; every article is reviewed and approved by a human editor before publication. Subscribe to the AI Worldwire newsletter. Last updated: 20 October 2026.