DeepSeek V4.1 Flash 1017: What The Pareto Upgrade And Open Search API Change For Builders
DeepSeek shipped a new checkpoint on 17 October 2026 and made two announcements at once. The first is DeepSeek-V4.1-Flash-1017, a minor update to its Flash model that the company describes as a Pareto improvement in performance and cost, with stronger benchmark results, faster task completion and shorter outputs than the September release. The second is that DeepSeek opened its Search and Fetch APIs to every developer at 0.75 dollars per 1,000 calls per endpoint.
The second announcement may matter more to anyone building agents. A coding model with a first-party web retrieval layer behind the same API key removes a vendor, a billing relationship and a chunk of glue code from the stack, and that is the pairing DeepSeek V4.1 Flash now offers.
This article covers what changed in DeepSeek V4.1 Flash, what the published benchmark numbers actually show, how to call the checkpoint without changing your code, what DeepSeek states about the Search and Fetch pricing, and where the evidence stops. Every figure below comes from DeepSeek's own release material or its model card, and the gaps are marked as gaps.
What DeepSeek actually shipped
DeepSeek V4.1 Flash 1017 is not a new architecture. It is an updated post-training checkpoint of DeepSeek V4.1 Flash, released on 17 October 2026 and served under the same model name, deepseek-flash. DeepSeek says the checkpoint delivers stronger benchmark performance, faster task completion and shorter outputs than the previous version, and it calls the combination a Pareto improvement in performance and cost.
Three facts about the release are worth separating from the marketing.
It ships under the same name, at what DeepSeek describes as the same position. The API model string does not change. DeepSeek's documentation states that deepseek-flash is now served by DeepSeek V4.1 Flash 1017 and that the calling method remains unchanged. Legacy names including deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, are served by the same checkpoint, and are billed at the Flash price.
The weights are open. DeepSeek published DeepSeek V4.1 Flash 1017 on Hugging Face and ModelScope under the MIT License, which permits commercial use. The model card includes recommended sampling parameters for self-hosted inference: temperature 1.0, top-p 1.0, a 1,000,000 token context window and a max tokens setting of 256,000 or more.
It is tuned for specific harnesses. DeepSeek says DeepSeek V4.1 Flash 1017 is optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends DeepSeek Harness v0.2.1-rc.1 or later. DeepSeek also states that the checkpoint uses reinforcement learning with self-feedback to improve end-to-end project development, with examples that range from game creation to 3D modeling.
That last point is the most interesting claim in the DeepSeek V4.1 Flash release for anyone who runs coding agents. A checkpoint tuned against particular harnesses is a different product from a general model, because the harness decides how tool calls are formatted, how much context is carried between steps and when the agent stops. If you use a different harness, part of the gain may not transfer, which is a reason to test rather than assume.
What the benchmark table shows
DeepSeek published a comparison across eleven benchmarks for DeepSeek V4.1 Flash 1017 on the Hugging Face model card. The figures below are DeepSeek's own, run in harnesses that DeepSeek names: Claude Code for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1.
| 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, or 44.0 on the text-only subset | 36.8, or 39.1 on the text-only subset | 43.5 | 42.0 on the text-only subset | 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 | Not disclosed | 79.9 | 84.0 |
| BabyVision with tools | 92.5 | 89.6 | 85.7 | Not disclosed | 88.9 | 94.1 |
| ZeroBench-main with tools | 51.0 | 49.0 | 41.0 | Not disclosed | 53.0 | 52.0 |
DeepSeek V4.1 Flash 1017 improves on the September checkpoint in every row, by between 2.0 points on ZeroBench-main and 11.1 points on Terminal-Bench 4.0. The strongest results for agent work are DeepSWE v1.1 at 77.0, Automation-Bench at 60.5 and Agents Last Exam at 35.4, where DeepSeek V4.1 Flash 1017 leads the reported field.
The losses matter as much as the wins. Claude Opus 5 remains ahead on Terminal-Bench at 51.8, on ProgramBench at 37.0, on NL2Repo-Bench at 75.3 and on HLE without tools at 56.3. In other words DeepSeek V4.1 Flash 1017 leads on long-horizon software engineering and business workflow tasks while a frontier model still leads on terminal-heavy execution, repository reconstruction, and raw knowledge.
That pattern is useful for routing decisions. If your workload is applying multi-file fixes inside a repository, DeepSWE is the closer proxy. If your workload is a terminal session with strict command constraints, Terminal-Bench is the closer proxy, and the gap there is nine and a half points in the other direction. A team that reads only the DeepSWE row will overestimate what DeepSeek V4.1 Flash 1017 will do for a shell-bound agent.
Two notes on reading the table. First, all of these are vendor-run results, and DeepSeek chose the harnesses and the comparison set. Second, the figures for Kimi K3, GLM 5.3, GPT 5.6 Sol and Claude Opus 5 are reported numbers rather than models re-run under identical conditions, which is normal for launch tables and also the reason they are not a substitute for your own evaluation.
How to call DeepSeek V4.1 Flash 1017
DeepSeek keeps the integration surface for DeepSeek V4.1 Flash stable, which is the practical advantage of a checkpoint release. The API speaks both the OpenAI and Anthropic formats, so an existing client only needs a base URL and a model string.
| Setting | Value |
|---|---|
| Base URL, OpenAI format | https://api.deepseek.com |
| Base URL, Anthropic format | https://api.deepseek.com/anthropic |
| Model string | deepseek-flash |
| Legacy aliases still accepted | deepseek-v4-flash and deepseek-v4-flash-vision-exp, both served by DeepSeek V4.1 Flash 1017 at the Flash price |
| Other model family on the same API | deepseek-v4-pro |
A minimal request is short. The snippet below leaves credential handling to your client configuration, which is the pattern to use in any script you keep, and it uses the thinking and reasoning effort controls that DeepSeek documents.
from openai import OpenAI
client = OpenAI(base_url="https://api.deepseek.com")
response = client.chat.completions.create(
model="deepseek-flash",
messages=[{"role": "user", "content": "Hello!"}],
extra_body={"thinking": {"type": "enabled"}, "reasoning_effort": "high"},
)
print(response.choices[0].message.content)DeepSeek also documents that the API works inside existing agent tools without code changes, naming Claude Code, GitHub Copilot and OpenCode as supported integrations, and it offers DeepSeek Harness as a developer preview for teams building their own harness.
For self-hosted deployment, the model card ships two pieces of tooling that matter more than they sound. There is no Jinja chat template in the release. Instead the encoding folder holds a self-contained Python reference implementation covering multi-turn conversations, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved images. Alongside it, deepseek-recipe provides Rust libraries with Python bindings that turn chat requests into the prompt format and parse the output back into responses. If you run DeepSeek V4.1 Flash locally and your prompts differ from the reference implementation, expect behavioural drift before you see any of the benchmark gains.
What DeepSeek says about Search and Fetch
The second half of the announcement is the one that changes architecture. DeepSeek opened the DeepSeek Search API and its Fetch counterpart to every developer, powered by an in-house search engine that the company says serves hundreds of millions of requests daily.
Three details are stated plainly in the release, and they are the ones to plan against.
- Price. Each endpoint costs 0.75 dollars per 1,000 calls.
- Free window. Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts on 31 October 2026 at 16:00 UTC.
- Documentation. The release points to the API guides for Search and Fetch, rather than documenting parameters in the announcement itself.
That free window is a deadline, not a discount. A team that builds retrieval into an agent during October and does not instrument call volume will discover the bill in November. The cost itself is small at the stated rate, since 1,000 successful calls cost 0.75 dollars per endpoint, but volume in an agent that searches per step adds up faster than most people expect. Model it before you ship.
There is a second reason the free window deserves attention. It applies to server-side calls through named harnesses, which means the same retrieval can be free or billable depending on how the agent is wired. If you proxy search yourself, you pay. If you let the harness do it, you do not, until the deadline.
For anyone comparing retrieval vendors, the useful test is not the price per call. It is whether the returned results improve your task success rate, and that requires running your own evaluation with your own queries. We made the same argument about retrieval and business meaning in our article on why agents ignore the governed semantic layer, and it applies here: a retrieval layer is only as good as the definitions and documents it can reach.
What this means for a working agent stack
Four practical moves follow from the release.
Pin the checkpoint in your evaluation, not just in production. Because deepseek-flash now points at DeepSeek V4.1 Flash 1017, an evaluation that ran in September is no longer describing the model you call today. Re-run your suite, and record which checkpoint produced which score. The harness matters too, and our guide to building production-grade agentic systems covers why the harness is part of the result rather than a neutral wrapper.
Measure output length and time to completion, not only pass rate. The central claim for DeepSeek V4.1 Flash 1017 is cheaper task completion through shorter outputs. If that holds in your workload, your cost per task falls without any change to per-token pricing. If it does not hold, the score gain may still be worth having, but the business case changes shape.
Use the benchmark rows that match your workload. DeepSWE for repository-level fixes, Terminal-Bench for shell-bound execution, Automation-Bench for business workflows, ProgramBench and NL2Repo for reconstruction. Do not average them into a single number, because the spread between rows is the useful information.
Treat the open weights as a real option. An MIT-licensed checkpoint at this capability level is unusual, and it means you can self-host for data-residency reasons or run a private evaluation without sending traffic to an API. The cost of that option is operational, since the model card expects you to handle encoding and inference yourself. Our analysis of long-horizon RL training explains why checkpoints trained for sustained work behave differently from models tuned for single answers, which is the capability DeepSeek is claiming here.
For orchestration, the release adds another credible option in the cost-sensitive tier, which is the same dynamic our survey of agent orchestration platforms tracks across vendors. And for teams running long coding sessions, the practical test remains the one we documented when we ran a coding agent overnight on a large codebase: what matters is whether the agent recovers from its own mistakes, not whether it wins a benchmark row.
Where the numbers stop
- Every benchmark figure in this article is DeepSeek's own, produced in harnesses DeepSeek selected. I did not reproduce any of them, and I did not find an independent replication of the 1017 checkpoint in the material I could read.
- The comparison columns for Kimi K3, GLM 5.3, GPT 5.6 Sol and Claude Opus 5 are reported figures rather than models re-run under identical conditions. Treat cross-model gaps as directional.
- DeepSeek's release describes the update as a Pareto improvement in performance and cost without publishing a per-token price table in the pages I could read. I have not stated specific token prices here, because I could not verify them from DeepSeek's own pricing documentation in this run, and an unverified price is worse than no price. Check the pricing page before you budget.
- The Search and Fetch details I could verify are the ones DeepSeek states: 0.75 dollars per 1,000 calls per endpoint, the free window through 31 October 2026 at 16:00 UTC for server-side calls in named harnesses, and an in-house engine. Parameter-level documentation sits in the API guides, which I could not read in this run.
- The 1,000,000 token context and 256,000 token output guidance for DeepSeek V4.1 Flash come from the model card as recommended settings for self-hosted inference. Actual limits on the hosted API may differ, and you should confirm them in the API reference.
FAQ
What is DeepSeek V4.1 Flash 1017?
DeepSeek V4.1 Flash 1017 is an updated post-training checkpoint of DeepSeek V4.1 Flash, released on 17 October 2026 and served under the model name deepseek-flash. DeepSeek describes it as a Pareto improvement in performance and cost, with stronger benchmark results, faster task completion and shorter outputs than the September release. The weights are published on Hugging Face and ModelScope under the MIT License.
Do I need to change my code to use it?
No. If you already call deepseek-flash, you are already calling DeepSeek V4.1 Flash 1017, because the model name now resolves to the new checkpoint and the calling method is unchanged. Legacy names such as deepseek-v4-flash are still accepted and route to the same checkpoint at the Flash price. Only self-hosted deployments need to pull new weights and follow the encoding and inference folders in the repository.
How much do the Search and Fetch APIs cost?
DeepSeek states 0.75 dollars per 1,000 calls for each endpoint. DeepSeek Search API and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing begins on 31 October 2026 at 16:00 UTC. The release note does not spell out how failed calls are billed, so confirm the billing rules in the API guides before you rely on them.
The short version
DeepSeek V4.1 Flash 1017 is a same-name checkpoint update that improves on the September model across eleven reported benchmarks, with the biggest jumps on Terminal-Bench and Automation-Bench and the strongest absolute score on DeepSWE at 77.0, while Claude Opus 5 still leads on terminal execution, repository reconstruction and raw knowledge. The weights are MIT licensed, the API string does not change, and the harnesses DeepSeek tuned for are named. In the same release DeepSeek opened Search and Fetch to all developers at 0.75 dollars per 1,000 calls per endpoint, with server-side calls free through 31 October 2026. The work to do is evaluation rather than integration: pin the checkpoint, re-run your suite in the harness you actually use, measure output length as well as pass rate, and calendar the end of the free retrieval window.
Sources
- DeepSeek, DeepSeek-V4.1-Flash-1017 release note, covering the Pareto claim, harness recommendations, self-feedback training, Search and Fetch pricing and the free window: https://api-docs.deepseek.com/news/news261017
- DeepSeek, API documentation, covering the OpenAI and Anthropic compatible base URLs, the deepseek-flash model name and the legacy aliases: https://api-docs.deepseek.com/updates
- DeepSeek on Hugging Face, DeepSeek-V4.1-Flash-1017 model card, covering the eleven benchmark rows, harness notes, MIT License and recommended sampling parameters: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017