DeepSeek V4.1-Flash-1017: What's New, API, Pricing and Benchmarks

DeepSeek V4.1-Flash-1017: What's New, API, Pricing and Benchmarks

AI Takeaway Teamon 2026-10-18

Last verified: October 18, 2026. Facts come from DeepSeek's official docs, changelog, Hugging Face page and X account linked below. Benchmark scores are vendor-reported unless marked otherwise. We have not run our own tests.

DeepSeek-V4.1-Flash-1017 is an update to DeepSeek's V4.1 Flash model, released on October 17, 2026. DeepSeek's changelog says V4.1 Flash "has been upgraded to DeepSeek-V4.1-Flash-1017," that the web and app models were updated too, and that "the API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model." The Hugging Face model card calls it "a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks." The weights are MIT-licensed. This guide covers what changed, how it compares with the earlier V4.1 Flash, how to call it, what it costs, and what early users say.

At a glanceDeepSeek-V4.1-Flash-1017
What it isUpdate to DeepSeek-V4.1-Flash (released September 10, 2026)
API model namedeepseek-flash
Context / max output1M tokens / up to 384K
Input typesText and images (vision supported)
Price (per 1M tokens, cache miss)$0.15 input / $0.60 output off-peak; $0.30 / $1.20 peak
License of open weightsMIT

Sources: Models & Pricing, Hugging Face, changelog.

Practical rule: if your code already calls deepseek-flash, you are on the 1017 model now. No code change and no price change.

What's new in V4.1-Flash-1017

DeepSeek's announcement on X describes it as "stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost," live on the website, app and API. The changelog adds:

  • Fewer tokens per task: the model's "task runtime and output length" were reduced, so tasks finish "at lower cost and higher speed."
  • Harness tuning: it was "specifically optimized" for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and DeepSeek recommends DeepSeek Harness 0.2.1-rc.1 or later.
  • Open weights: published on Hugging Face under the MIT license, with recommended sampling of temperature 1.0 and top_p 1.0, a 1M-token context window and max_tokens of at least 256K (model card).
  • Search & Fetch APIs: released the same day as standalone APIs on DeepSeek's in-house search engine, priced at $0.75 per 1,000 successful requests. Server-side tool calls through DeepSeek Harness, Claude Code or Codex are free until billing starts at 00:00 Beijing time on November 1, 2026 (pricing page).

V4.1-Flash-1017 vs the earlier V4.1 Flash

All numbers below are vendor-reported, from the Hugging Face model card and the changelog. DeepSeek's own comparison also lists rival models; the table keeps only the two DeepSeek versions.

BenchmarkV4.1-Flash-1017V4.1-Flash (Sep 10)
HLE, no tools (text-only subset in brackets)42.6 (44.0)36.8 (39.1)
Terminal-Bench 4.0 (Claude Code harness)42.331.2
DeepSWE v1.1 (Mini-SWE-V2 harness)77.074.2
ProgramBench29.120.3
NL2Repo-Bench68.265.4
Automation-Bench v1.0.660.554.8
Agents' Last Exam35.431.8
Apex Bench v1.172.164.6
Chartography (with tools)83.078.9
BabyVision (with tools)92.589.6
ZeroBench-main (with tools, Pass@5)51.049.0

On the model card, DeepSeek also compares the new model with Kimi K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Its own DeepSWE v1.1 figure is 77.0 for 1017 against 73.0 for GPT-5.6 Sol and 74.0 for Opus 5, while Opus 5 stays ahead on HLE and Terminal-Bench 4.0 (56.3 and 51.8, against 42.6 and 42.3). DeepSeek calls this "a minor update," and several Hacker News commenters joked about that wording next to such large jumps.

Treat vendor tables as a starting point. At the time of writing, no independent leaderboard result for 1017 had been published in the sources we checked.

How to use the API

Per the API docs, the API is compatible with the OpenAI and Anthropic formats:

  • OpenAI-format base URL: https://api.deepseek.com
  • Anthropic-format base URL: https://api.deepseek.com/anthropic
  • Model name: deepseek-flash (served by DeepSeek-V4.1-Flash-1017). The other current model is deepseek-v4-pro.
  • The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are "still accepted," but those models are retired and requests are served by 1017 at the Flash price.
  • A Responses API endpoint (POST /responses) is also available, and deepseek-flash accepts image inputs there.

Here is the official Python example from the docs (install the OpenAI SDK first with pip3 install openai):

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ.get('DEEPSEEK_API_KEY'),
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[
        {"role": "system", "content": "You are a helpful assistant"},
        {"role": "user", "content": "Hello"},
    ],
    stream=False,
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)

print(response.choices[0].message.content)

And the curl version:

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
        "model": "deepseek-flash",
        "messages": [
          {"role": "system", "content": "You are a helpful assistant."},
          {"role": "user", "content": "Hello!"}
        ],
        "thinking": {"type": "enabled"},
        "reasoning_effort": "high",
        "stream": false
      }'

The pricing page says the model supports both thinking (the default) and non-thinking modes, JSON output, tool calls, the Responses API, the Anthropic API and vision. FIM completion is non-thinking mode only. For the thinking effort levels (low, high, max), see DeepSeek's Thinking Mode guide.

Pricing

Prices are per 1M tokens, from the official pricing page:

Off-peakPeak
Input, cache hit$0.003$0.006
Input, cache miss$0.15$0.30
Output$0.60$1.20

Off-peak rates are half the peak rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; everything else is off-peak. The listed concurrency limit for deepseek-flash is 2,500. For comparison, deepseek-v4-pro costs $0.66 input and $1.98 output off-peak on the same page.

Early reviews and reactions

The model had been out for about a day when we checked, so these are first impressions, not controlled tests.

Hacker News. The launch thread had 846 points and 645 comments. Typical remarks:

  • "Those benchmarks can't possibly be real, right? Right?" (freakynit)
  • "Those are crazy numbers for a Flash model." (nik736)
  • "Looks very impressive especially considering the price and speed. Let's wait for AA to confirm how well their benchmarks hold up." (waximabbax)
  • "...it's noticeable smarter, beats GLM" (floooof), from a user who says they had been using it in OpenCode for a couple of days.
  • "I do wish that we will see some smaller 'flash' models again though - running this one locally is tough (very costly)." (jononor)

X.

  • Chubby (@kimmonismus), repeating DeepSeek's comparison: "DeepSeek V4.1 Flash‑1017 beats GPT‑5.6 Sol and Opus 5 on selected coding and agent benchmarks, at much lower token prices. Super duper impressive!" (post). Note "selected" benchmarks, and these are vendor numbers.
  • Leo (@leo114119) ran a small real coding task, in Chinese, and said the improvement in real work is "very noticeable" while the task was too easy to stretch the model (post).
  • Ollama says the model is available on its cloud with US-based hosting and zero data retention (post).

The limits are clear: the enthusiastic posts mostly restate DeepSeek's benchmark chart, and the hands-on posts are single tasks. For a longer hands-on test, the independent video below runs six app-building tasks:

Watch on YouTube: "DeepSeek V4.1 Flash 1017: I Tested It on 6 Real Apps for $1.97 (KingBench 4)" by AISeeKing

Third-party video by AISeeKing, an independent creator. AI Takeaway did not make this video and is not affiliated with its creator.

No time to watch the whole review? AI Takeaway is a tool that works on any YouTube video, including third-party reviews like this one. Paste the link to get the key points in minutes, jump to the original timestamps, and ask follow-up questions; every answer cites the video. You can also browse the Hub for more AI and tech interviews.

Who should use it

  • Teams already on deepseek-flash: you get the upgrade automatically at the same price.
  • Agentic coding on a budget: DeepSeek's pitch is lower cost per task through shorter outputs. Run your own tasks and compare cost per finished task, not just per-token price.
  • Image-in workflows: vision is supported on deepseek-flash but not on deepseek-v4-pro.
  • Self-hosters: the weights are MIT-licensed, but local running is demanding; one Hacker News user called it "very costly."
  • Probably not for: anyone who needs independent benchmark confirmation first, or anyone who cannot send data to DeepSeek's hosted API and has no way to host the weights.

How it compares with other recent models (briefly)

DeepSeek's chart pits it against much pricier models, and early commenters repeat that framing. For other recent launches in the same race, see our guides to GPT-6.1 Sol vs GPT-6 Astra, Gemini 4 Argon, Meta's Muse Wonder and Grok Voice 3.0.

Getting started checklist

  1. Create a DeepSeek API key and set DEEPSEEK_API_KEY.
  2. Point an OpenAI-compatible client at https://api.deepseek.com with model="deepseek-flash".
  3. Choose a thinking effort (reasoning_effort) that fits your task.
  4. Run 20–50 of your own tasks and track pass rate, tokens and cost.
  5. Check Models & Pricing for peak and off-peak hours before scheduling batch jobs.

Keep up without the watch time

New models land every week. If you would rather read than watch, the AI Takeaway video summary hub has free curated summaries of AI and tech interviews, for example:

To keep up with ongoing AI news, check the Hub regularly. You can also install the AI Takeaway Chrome extension to get key points right on YouTube, or submit any YouTube video yourself on aitakeaway.net.

For more model guides and comparisons, browse all AI Takeaway articles.

FAQ

What is DeepSeek-V4.1-Flash-1017?

An October 17, 2026 update to DeepSeek-V4.1-Flash with better benchmark scores, faster task completion and shorter outputs, according to DeepSeek. It is served through the API as deepseek-flash.

How much does it cost?

$0.15 per 1M input tokens (cache miss) and $0.60 per 1M output tokens off-peak; $0.30 and $1.20 at peak. The price is unchanged from the earlier V4.1 Flash.

What changed compared with V4.1 Flash?

On DeepSeek's own benchmarks, DeepSWE v1.1 rose from 74.2 to 77.0 and Terminal-Bench 4.0 from 31.2 to 42.3, with shorter outputs and faster tasks.

Do I need to change my code?

No. Keep using deepseek-flash. If you still use deepseek-v4-flash, it is routed to the new model.

Is it open source?

The weights are on Hugging Face under the MIT license.

Does it support images?

Yes. deepseek-flash supports vision input; deepseek-v4-pro does not.

Sources

DeepSeek V4.1-Flash-1017: What's New, API, Pricing and Benchmarks