DeepSeek launches V4.1-Flash-1017: one-click coding brings ideas to life, closing in on GPT-5.6
TL;DRDeepSeek V4.1-Flash-1017 is live on the web, mobile app, and API. DeepSeek reports stronger coding and agent performance, shorter task runtimes, and less output, with API prices unchanged. Its release demos show complete projects built from brief requests. Developers can keep using deepseek-flash and should update DeepSeek Harness to 0.2.1-rc.1 or later.
By DeepSeek V4 Hub
DeepSeek released V4.1-Flash-1017 on October 17, 2026, updating the model behind its official website, app, and Flash API. The company describes better benchmark results alongside lower task costs, then makes its case through four projects: a 3D Forbidden City, a browser racing game, a Godot recreation of Portal 2, and a changing landscape that follows human history. [1]
DeepSeek calls this "wish-driven coding": describe the project you want, provide a small amount of reference material, and let the agent handle more of the implementation. The company says it produced the demos in DeepSeek Harness with very few interaction rounds. [1]
How much better is V4.1-Flash-1017?
DeepSeek reports improvements across reasoning, coding, agent tasks, and visual benchmarks. Terminal-Bench 4.0 rises from 31.2 to 42.3, an increase of 11.1 percentage points. ProgramBench moves from 20.3 to 29.1, while DeepSWE v1.1 improves from 74.2 to 77.0. [2]
Scroll horizontally to compare all six models.
| Benchmark | DeepSeek V4.1-Flash-1017 | DeepSeek V4.1-Flash | Kimi K3 | GLM 5.3 | GPT-5.6 Sol | Claude Opus 5 |
|---|---|---|---|---|---|---|
| HLE without tools | 42.6 / 44.0* | 36.8 / 39.1* | 43.5 | 42.0* | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents' Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| Chartography (with tools) | 83.0 | 78.9 | 68.1 | n/a | 79.9 | 84.0 |
| BabyVision (with tools) | 92.5 | 89.6 | 85.7 | n/a | 88.9 | 94.1 |
| ZeroBench-main (with tools) | 51.0 | 49.0 | 41.0 | n/a | 53.0 | 52.0 |
Source: DeepSeek's official model card. Bold indicates the highest score in each row. * HLE text-only subset. "n/a" means the source does not report a result. [3]
Get started with the latest DeepSeek V4.1 Flash 1017 in one click. Explore our Coding Plan.
The evaluation setup matters. DeepSeek used Claude Code for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1. The starred HLE figures cover only its text subset; the unstarred figures cover the broader evaluation. ZeroBench-main uses Pass@5, so its score should not be read as a single-attempt success rate. These are DeepSeek's reported results. [2] [3]
Is DeepSeek closing in on GPT-5.6?
On this particular benchmark table, V4.1-Flash-1017 is already ahead of GPT-5.6 Sol on nine of the eleven rows. DeepSWE is 77.0 versus 73.0, and Terminal-Bench 4.0 is 42.3 versus 39.9. GPT-5.6 Sol remains ahead on HLE without tools and ZeroBench-main. [3]
More importantly, these results come from DeepSeek's Flash model. In our view, it feels considerably faster than GPT-5.6 Sol, making its coding and agent benchmark results even more compelling.
V4.1-Flash-1017 is competitive on these coding and agent evaluations. A different repository or prompt may produce a different ranking. The table also shows Claude Opus 5 leading several tests, including HLE, Terminal-Bench 4.0, and ProgramBench. [3]
Does better performance come with higher costs?
DeepSeek says it reduced task runtime and output length while improving results. Its performance-versus-cost chart uses DeepSWE v1.1 with the mini-swe-v2 framework and plots Pass@1 against average task cost in renminbi. The horizontal axis is logarithmic. [1]
Chart source notes
The source figure cites the DeepSWE v1.1 leaderboard, the mini-swe-agent evaluation framework, and data dated September 22, 2026. It attributes the GPT-6 Sol, GPT-6.1 Sol, GPT-6 Astra, and GPT-6 Luna points to OpenAI's official DeepSWE v1.1 figure. USD costs, including historical V4 Flash costs, use the figure's conversion of US$1 = CNY 6.6862, dated October 16, 2026. These are the source figure's methods and references.
The company describes the result as a better performance-cost frontier across multiple harnesses: users can complete tasks with stronger results and less spending. The chart's 77.0% point is the maximum-effort DeepSWE result, rather than a success rate for arbitrary coding requests. [1]
The model's API prices remain unchanged. Using fewer output tokens can lower a task's cost at the same token rate, while shorter runtimes reduce the wait for a result. DeepSeek says this release received specific optimizations for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy, and other platforms. [2]
What can it build from a brief request?
DeepSeek attributes the model's improved ability to build complete projects from scratch to reinforcement learning based on self-feedback. The release focuses on game development and 3D modeling, where an agent has to turn a short description into a working project. [1]
The four official examples below used limited prompts and reference material inside DeepSeek Harness. They show the results DeepSeek chose to publish; the announcement does not promise that every request will produce a finished project in a single attempt. [1]
The Forbidden City in 3D
A racing game in the browser
Portal 2, recreated in Godot
Valley of Ages
What do the new Search and Fetch APIs do?
DeepSeek has also opened its search infrastructure to developers. Search returns structured web results for a query; Fetch retrieves a page's content as Markdown. Both APIs use the same bearer authentication as the DeepSeek API. [4]
DeepSeek says its search engine already processes hundreds of millions of requests each day and indexes hundreds of billions of pages. It supports multilingual search and filters SEO spam, GEO manipulation, and low-quality AI-generated content. Those descriptions of scale and filtering come from DeepSeek. [1]
The Chinese announcement lists each API at RMB 5 per 1,000 requests. The English documentation lists each at US$0.75 per 1,000 successful requests. Search charges require a nonempty result list; Fetch charges require returned page content. Empty searches, failed requests, and timeouts do not incur a charge. [1] [4] [5]
Server-side Search and Fetch calls made through DeepSeek Harness, Claude Code, or Codex are temporarily free. DeepSeek says per-request billing for these calls begins on November 1, 2026, at 00:00 Beijing time, which is October 31 at 16:00 UTC. [4]
Where can you download the model weights?
DeepSeek has released V4.1-Flash-1017 weights on Hugging Face and ModelScope. The official model card licenses the repository and weights under MIT and provides links to inference instructions. [1] [3]
How should you try the update?
- Keep
deepseek-flashas your API model name. The October 17 update does not require a new dated model ID. [2] - Upgrade DeepSeek Harness to
0.2.1-rc.1or later before comparing results. [2] - Run a familiar task with the same prompt and effort setting. Compare the finished result, elapsed time, and token use with your previous run.
- For a project request, state the required behavior and give the agent a way to check it. Review and run the result before using it.
Frequently asked questions
Do I need to change my API configuration?
Keep using deepseek-flash. DeepSeek says the calling method and pricing are unchanged. [2]
Does "one-click coding" mean a guaranteed finished app?
The release demonstrates projects made with brief prompts and very few interaction rounds. The examples support trying more ambitious requests, while their results remain specific to those demonstrations. [1]
Are Search and Fetch included in model token prices?
Search and Fetch are separate APIs billed per successful request. The temporary exemption for integrated server-side calls has the end date described above. [4] [5]