DeepSeek V4.1 Flash 1017 review roundup: coding scores above Opus 5 at cents per task
DeepSeek V4.1 Flash moved to its 1017 build on October 17, 2026: same price, open weights. Vendor scores, independent tests and Bilibili creators agree it builds new projects well for very little; reliability on existing code is unproven.
In short
- DeepSeek V4.1 Flash launched in September 2026 and moved to its 1017 build on October 17: the same price and model name, weights open under the MIT license, image input supported.
- DeepSeek reports 77 on DeepSWE v1.1, above Claude Opus 5's 74, at about ¥1.3 per task. On the independent Bug Hunt Bench it fixed 34 bugs for $0.90; Claude Opus 5.5 fixed 41.7 for $58.50.
- Bilibili creators agree that it builds from scratch quickly and cheaply. The weak spots are Chinese creative writing, basic image reading, and the need to adapt it separately to each agent framework.
- deepseek-v4.1-flash-1017 was the second most used model on MoeRouter over the last 7 days. Use it as the workhorse for daily coding and batch jobs, and escalate the hard problems.
On this page
What DeepSeek V4.1 Flash 1017 is#
DeepSeek V4.1 Flash is the main model DeepSeek released in September 2026. On October 17 it moved to the 1017 build: the website, the app and the API switched together, the API model name and prices stayed as they were, and the weights were opened under the MIT license[1]. DeepSeek's model card calls it a "minor update" with "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks", plus shorter outputs and faster task completion[2].
| DeepSeek V4.1 Flash 1017 | |
|---|---|
| List price per 1M tokens | $0.30 input and $1.20 output at peak hours, half that off-peak |
| Input and output | Text and images in, text out |
| Context and output | 1M-token context, up to 384K output tokens |
| Weights | Open, MIT license |
| Model names on MoeRouter | deepseek-v4.1-flash-1017 (pinned to the October 17 build) and deepseek-v4.1-flash (follows the current build) |
| Endpoints on MoeRouter | Chat Completions, Responses, Messages, Gemini |
Prices are as Atomic Agent reports them, and specifications as Kingy AI compiled them from DeepSeek's documentation[2][3].
What it costs on MoeRouter#
| Model | Group | Input | Output | Cache read |
|---|---|---|---|---|
deepseek-v4.1-flash-1017 | default | CN¥0.1≈ US$0.015 | CN¥0.4≈ US$0.06 | CN¥0.002≈ US$0.0003 |
tencentcloud | CN¥1.2≈ US$0.18 | CN¥4.8≈ US$0.72 | CN¥0.024≈ US$0.0036 | |
deepseek-v4.1-flash | default | CN¥0.1≈ US$0.015 | CN¥0.4≈ US$0.06 | CN¥0.002≈ US$0.0003 |
Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. The same model takes other routes in another group and costs differently there; a key's group is chosen when the key is created. About groups Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().
deepseek-v4.1-flash-1017 costs very differently in the default and the tencentcloud groups, so check the group before creating a key. Which name to use: the one ending in 1017 if a project needs a fixed build, deepseek-v4.1-flash otherwise.
Benchmarks#
Reported by DeepSeek#
By the release notes and the model card, 1017 beats the previous build on all 11 benchmarks DeepSeek published[1]:
| Benchmark | 1017 | Previous build | For comparison |
|---|---|---|---|
| DeepSWE v1.1 | 77 | 74.2 | Claude Opus 5 74, GPT-5.6 Sol 73 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | GPT-5.6 Sol 39.9 |
| ProgramBench | 29.1 | 20.3 | — |
A DeepSWE task costs about ¥1.3 on average, against ¥3.3 on the previous build[10]. Note what DeepSeek compares with: Claude Opus 5 and GPT-5.6 Sol, not the newer Opus 5.5 or the GPT-6 family. And of the 11 comparisons, Claude Opus 5 still leads 6, including HLE, while 1017 leads 4[1].
Measured by third parties#
- Bug Hunt Bench seeds 105 bugs in two real codebases. 1017 at maximum effort fixed 34 and the previous build 21.7; Claude Opus 5.5 fixed 41.7 for $58.50 where 1017 spent $0.90[1].
- Atomic Agent's coding-agent test ran the same task in the same agent three times per model. 1017 passed all 96 hidden tests in 27 seconds and 14 steps on average for $0.016 per task; Claude Sonnet 5.5 passed them all too, in 49 seconds for $0.183. That is "about 11 times cheaper" and "almost twice as fast". The author adds that on this task 1017 was neither faster nor cheaper than the September build; the gains DeepSeek reports show on harder benchmarks[2].
- Shorter outputs. An inference provider ran the same set of maths-competition problems on both builds: the median output fell from 1,681 tokens to 1,124 with about the same number solved[1].
Hands-on: what Bilibili creators found#
karminski-牙医 is the most enthusiastic and also names the most specific problem. His description calls it a model you must not miss, with back-end agentic coding and front-end work close to the "wish-style programming" DeepSeek advertises. The problem is that its agent behaviour seems to have been trained to be stingy: when an action has a cost, it tends not to take it. That is entirely unlike the previous Flash build, so he warns anyone plugging it into their own agent framework to adapt it separately or expect it to go wrong (our translation)[4].
碳基生物退役指南 ran 11 tests and is more measured: a marked improvement in code and complex engineering delivery with a third fewer tokens, but a step back in Chinese creative writing and basic image reading; overall it neither fully surpasses the old build nor catches flagships such as GPT-6 Astra, yet it is worth trying for people who mainly write code[5].
AI超元域 ran seven development tasks in DeepSeek Harness at high effort, among them an art-history animation, the Rainbow Bridge from Along the River During the Qingming Festival, a Godot 4 tank game and a native iOS app: 600 million tokens for ¥30. The bridge took about two hours, the tank game 30 minutes and a vocabulary app about 50 minutes[6].
kate人不错 used 460 million tokens, about ¥24 at off-peak prices, on fiction writing and review, Blender modelling, a brick-building demo and a product explainer, showing both what one prompt produced and the problems in detail, picture and sound[7].
花叔v made nine cases with 500 million tokens (animation, games, 3D), and states that he gave feedback and made adjustments along the way: none was a single shot[8].
程序员阿江-Relakkes tested two cases during the September beta. An online spreadsheet built in DeepSeek Harness came out about level with GPT-6 and Claude Fable 5.1; on a complex robot-scheduling system GPT-6 and Claude Fable 5.1 were clearly better. He also found the same model did slightly worse in Claude Code than in DeepSeek Harness[9].
Weak spots and open questions#
- Maintaining existing code is unproven. SMZDM's roundup notes developers reporting that, when continuing an existing game demo, 1017 deleted interactions that were there and broke one thing while fixing another. These are individual reports, not evidence of a general regression, but they mark the distance between generating code and maintaining it[1].
- Still behind the flagships on complex systems. On a complex scheduling system, 3D world knowledge and SVG taste, several creators find it short of GPT-6 Astra and the Claude flagships[5][9].
- A new framework means adapting again. It was tuned for DeepSeek's own Harness; in another agent framework it may behave differently[1][4].
- Teams should run a regression first. An unchanged model name only means the entry point is the same; run frozen real tickets against acceptance criteria set in advance before sending it volume[1].
How to use it#
Good for
- Prototypes from scratch and small single-purpose projects
- Daily coding and high-volume batch work, at cents per task
- Simple tasks that need to read an image, such as screenshots and charts
Less suited to
- Long-term maintenance of a large existing project (try it on a small part first)
- Design of complex multi-module systems
- Chinese creative writing
A thrifty set-up: DeepSeek V4.1 Flash by default, and GPT-6.1 Sol or Claude Opus 5.5 for the tasks that fail or are plainly hard.
How to call it#
On MoeRouter it answers on all four endpoints, so the OpenAI SDK, Claude Code, Codex and Gemini CLI can all use it. With the OpenAI SDK:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://moerouter.com/v1",
api_key=os.environ["MOEROUTER_API_KEY"],
)
reply = client.chat.completions.create(
model="deepseek-v4.1-flash",
messages=[{"role": "user", "content": "Write a paginated to-do API in FastAPI, with pytest cases"}],
)
print(reply.choices[0].message.content)
In Claude Code:
export ANTHROPIC_BASE_URL="https://moerouter.com"
export ANTHROPIC_AUTH_TOKEN="$MOEROUTER_API_KEY"
unset ANTHROPIC_API_KEY
claude --model deepseek-v4.1-flash
The docs are in Chinese for now.
Frequently asked questions#
What is the difference between deepseek-v4.1-flash and deepseek-v4.1-flash-1017?#
deepseek-v4.1-flash-1017 is pinned to the October 17 build, for projects that need a fixed version; deepseek-v4.1-flash follows the current build. Use the latter unless you need to pin.
How much does the DeepSeek V4.1 Flash API cost?#
List price per 1M tokens is $0.30 input and $1.20 output at peak hours and half that off-peak. MoeRouter's reference prices are in the table above; accounts are charged in Chinese yuan.
Is DeepSeek V4.1 Flash really better than Claude Opus?#
On DeepSeek's own DeepSWE v1.1 figures it is above Claude Opus 5 (77 against 74), but that is not the newer Opus 5.5. On the independent Bug Hunt Bench, Opus 5.5 fixed more bugs (41.7 against 34) at 65 times the cost. Its advantage is cost, not the ceiling.
Does DeepSeek V4.1 Flash accept images?#
Yes. It takes image input natively and can read screenshots and charts. One creator's tests found basic image reading weaker in the 1017 build than before, so test first if image reading matters.
Can I use DeepSeek V4.1 Flash in Claude Code or Codex?#
Yes. On MoeRouter it supports both the Messages and the Responses endpoints; set the model to deepseek-v4.1-flash. Reviewers found it behaves differently from one agent framework to another, so run a few of your own tasks after switching.
Why is it second by usage on MoeRouter?#
Over the 7 days to November 18, 2026, deepseek-v4.1-flash-1017 used more tokens than any model except GPT-6.1 Sol. Its price is very low, which suits high-volume work, and high-volume work uses many tokens.
Sources
The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.
- DeepSeek新模型V4.1-Flash-1017实测汇总:编程跑分超Opus 5、单任务低至1.3元,但别急着把存量项目交给它 (in Chinese)
- DeepSeek V4.1 Flash 1017: $0.016 per Coding Task
- DeepSeek V4.1 Flash 1017: Specs, Benchmarks, Pricing and Frontier Rivals
- DeepSeek 1017 变抠门了??? (video, in Chinese)
- DeepSeek V4.1 Flash(1017新版)实测! (video, in Chinese)
- 🚀6亿Token只花了30元!我用DeepSeek-V4.1-Flash-1017连做7项开发测试,表现到底如何?在DeepSeek Harness中实测 (video, in Chinese)
- DeepSeek V4.1 Flash 1017 实测:4.6 亿 Token 花了约 24 元,效果如何? (video, in Chinese)
- 5亿token实测!DeepSeek-V4.1-flash-1017做动画能力超GPT-6! (video, in Chinese)
- DeepSeek V4.1 Flash突然内测,实测两个Case下来有点猛啊 (video, in Chinese)
- DeepSeek V4.1 Flash更新到1017:763B参数、成本降2/3、11项基准全涨、API自带搜索,这还叫Flash? (video, in Chinese)
Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.