DeepSeek-V4.1-Flash-1017 — an update to the DeepSeek V4.1 Flash model that the company released on October 17, 2026. According to DeepSeek, the model is stronger on agentic and coding benchmarks, finishes tasks faster, and produces shorter responses. API pricing as of October 18 starts at $0.15 per 1 million input tokens off-peak and goes up to $1.20 per 1 million output tokens during peak hours. Along with the model, Search API and Fetch API are now available — $0.75 per 1,000 successful requests. What matters for businesses: under the old names deepseek-flash and deepseek-v4-flash the new version now responds, so working workflows should be retested.
Prepared by AIrassvet October 18, 2026 based on DeepSeek’s official pages: the release announcement, pricing page, Search and Fetch API guide, and the model card on Hugging Face. It is intended for IT leaders, product teams, and business owners who already use DeepSeek or are choosing a low-cost model for AI agents. We did not test the model on our own tasks: all benchmark results are claimed by DeepSeek, and the calculations, checklists, and selection table are the editorial team’s methodology. AIrassvet is not affiliated with DeepSeek; the section at the end of the article describes the publisher’s services.
Contents
- At a glance
- What DeepSeek released on October 17
- What changed in the benchmarks
- How much DeepSeek V4.1 Flash API costs
- Example calculation for an AI assistant
- Search API and Fetch API: what they are and what they cost
- The model was updated under the same name: what to check
- API or your own servers
- Data and legal requirements
- When to choose V4.1 Flash, and when not to
- FAQ
- Primary sources and verification method
At a glance
- What was released. DeepSeek-V4.1-Flash-1017 is an intermediate update to DeepSeek-V4.1-Flash from September 10, 2026. The weights are open under the MIT License.
- What improved. According to DeepSeek’s table, Terminal-Bench 4.0 rose from 31.2 to 42.3, Automation-Bench from 54.8 to 60.5, and DeepSWE v1.1 from 74.2 to 77.0.
- What it costs. During peak hours: $0.30 per 1 million input tokens without cache and $1.20 per 1 million output tokens. Off-peak — half the price.
- Peak hours in Moscow time. From 04:00 to 07:00 and from 09:00 to 13:00 on weekdays — that is the morning of the workday for Russian companies.
- New. Search API and Fetch API for agents and RAG — $0.75 per 1,000 successful requests; the limit is 5 requests per second per account.
- What to do. Rerun your test scenarios: the model name in the code stayed the same, but the version behind it changed.
What DeepSeek released on October 17
DeepSeek-V4.1-Flash-1017 is the updated version of the DeepSeek-V4.1-Flash model. The “1017” in the name indicates the release date — October 17. DeepSeek calls the release a “minor update,” but says it brings a noticeable improvement in reasoning, coding, agentic workflows, and image analysis tasks.
What is included in the release, according to DeepSeek’s announcement and the model card on Hugging Face:
- the model is available on the website, in the app, and through the API under the name
deepseek-flash; - according to the company, the model finishes tasks faster and gives shorter responses — meaning fewer output tokens are needed per task;
- the model is optimized for the DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy wrappers; for DeepSeek Harness, version v0.2.1-rc.1 or newer is recommended;
- Search API and Fetch API are now available on DeepSeek’s own search engine;
- the weights are published on Hugging Face and ModelScope under the MIT License.
Wrapper (harness) is the software layer around a model: it manages context, calls tools, and checks the result. For more on why this affects agent performance, see the article on accepting an office AI agent on your own tasks.
Release context. On September 10, 2026, DeepSeek introduced V4.1-Flash — a MoE model with 552 billion parameters, of which 8 billion are active for input processing and 16 billion for generation. The company said that the KV cache in the new architecture uses 4 times less GPU memory and 8 times less SSD space than the previous generation. Version 1017 continues this line. We wrote about the previous generation in the overview DeepSeek V4 Flash: local deployment and API.
What changed in the benchmarks
All figures below are taken from the table in the model card and are claimed by DeepSeek itself. We did not find an independent verification as of October 18. For Terminal-Bench 4.0, the company used the Claude Code wrapper; for DeepSWE v1.1, it used mini-swe-v2.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash (September) | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 (terminal work) | 42.3 | 31.2 | 39.9 | 51.8 |
| DeepSWE v1.1 (code fixing) | 77.0 | 74.2 | 73.0 | 74.0 |
| Automation-Bench v1.0.6 (automation) | 60.5 | 54.8 | 45.8 | 50.3 |
| Agents' Last Exam (agent tasks) | 35.4 | 31.8 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 54.3 | 65.8 |
| HLE without tools | 42.6 | 36.8 | 44.5 | 56.3 |
| BabyVision with tools (vision) | 92.5 | 89.6 | 88.9 | 94.1 |
How to read the table:
- Growth compared with the September version is shown in every row. The largest gain is on Terminal-Bench 4.0: +11.1 points.
- On agentic and automation benchmarks DeepSeek shows its model ahead of GPT-5.6 Sol and Opus 5. On Terminal-Bench 4.0, Opus 5 remains ahead, while on the tool-free reasoning benchmark (HLE), both Opus 5 and GPT-5.6 Sol lead.
- A benchmark is not the same as your task. These are results on developer benchmark sets. We covered how to build your own evaluation in the article "Your Own Benchmark for Choosing an LLM".
How much the DeepSeek V4.1 Flash API costs
Prices are from the Models & Pricing page as of October 18, 2026, in U.S. dollars per 1 million tokens. The model deepseek-flash is currently served by DeepSeek-V4.1-Flash-1017.
| Pricing tier | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| Peak hours | $0.006 | $0.30 | $1.20 |
| Off-peak | $0.003 | $0.15 | $0.60 |
Other parameters from the same page:
- context — 1 million tokens, maximum response — 384,000 tokens;
- modes — with reasoning (default) and without it;
- compatibility — OpenAI format (
https://api.deepseek.com), Anthropic format (https://api.deepseek.com/anthropic) and Responses API; - features — JSON output, tool calling, image handling;
- parallel request limit — 2500 for
deepseek-flash.
Peak hours converted to Moscow time (UTC+3): from 04:00 to 07:00 and from 09:00 to 13:00 Monday through Friday. This is the editorial conversion of the DeepSeek intervals 01:00–04:00 and 06:00–10:00 UTC. Weekends and Chinese public holidays are entirely counted as off-peak. For a Russian company, this means: morning weekday hours are billed at full price, while evening and night hours are twice as cheap.
Practical takeaway: batch jobs—document labeling, archive summarization, overnight reports—are best run after 1:00 p.m. Moscow time or on weekends.
Separately, about the deepseek-v4-pro. In its September announcement, DeepSeek said requests to it would be routed to V4.1-Flash until V4.1-Pro was released. On the pricing page dated October 18, deepseek-v4-pro is listed separately again—as DeepSeek-V4-Pro-0813 with its own pricing. If you use this name, verify the actual version in the API responses. Our breakdown of the previous version is in the article DeepSeek V4 Pro: what's new and how much it costs.
Example calculation for an AI assistant
The calculation below is an editorial example, not a benchmark. Let’s take a support assistant that handles 2,000 requests per day. Each request includes:
- 4,000 input tokens that repeat and get cached (system prompt and help content);
- 2,000 new input tokens (customer question and conversation history);
- 500 output tokens (response);
- 1 Search API request.
| Cost item | Daily volume | Peak hours | Off-peak |
|---|---|---|---|
| Cached input | 8 million tokens | $0.048 | $0.024 |
| Uncached input | 4 million tokens | $1.20 | $0.60 |
| Output | 1 million tokens | $1.20 | $0.60 |
| Search API | 2,000 requests | $1.50 | $1.50 |
| Total per day | — | $3.95 | $2.72 |
What the calculation shows:
- search costs about as much as the model itself, and off-peak it is even more expensive ($1.50 versus $1.22 per day in this example). If the agent searches the web at every step, count Search API requests as a separate budget line;
- the cache is almost free. It pays to keep the stable part of the instructions at the beginning of the prompt so it gets cached;
- tokens are counted differently across languages. Russian text is usually split into more tokens than English — more on that in the article “Russian Text Tokenization and LLM Cost”.
The pricing page does not say whether the off-peak discount applies to Search and Fetch, so in this calculation their price is the same in both columns.
Search API and Fetch API: what they are and how much they cost
Search API searches web pages based on a query and returns a structured list of results. Fetch API downloads a single page and returns its contents in Markdown. Both run on the DeepSeek search engine and are designed for agents and RAG systems that need fresh web data.
What you need to know from the Search & Fetch guide:
| Parameter | Search API | Fetch API |
|---|---|---|
| Endpoint | POST /v1/tools/search |
POST /v1/tools/fetch |
| What to send | query up to 128 characters, number of results from 1 to 50 | page URL up to 8192 characters |
| What it returns | title, link, short and long description, date, relevance score, and source authority level | title and page text in Markdown |
| Price | $0.75 per 1,000 successful requests | $0.75 per 1,000 successful requests |
| Limit | 5 requests per second per account | 5 requests per second per account and a per-domain limit |
Details that affect budget and architecture:
- you pay only for successful requests. Empty search results, errors, and timeouts are not billed; the number of results in the response does not affect the price;
- search operators are supported:
site:, site and word exclusions, exact phrase, languagelang:, datesafter:andbefore:; - each result has a score
rerank_scorefrom 0 to 3 (1.5 and above means a quality result) and anauthorityfrom 0 to 4. Based on these, the agent can discard weak sources; - a limit of 5 requests per second applies to the entire account, not to a key. For a service with hundreds of concurrent users, a queue is needed;
- free until November 1. Search calls from DeepSeek Harness, Claude Code, and Codex are not yet billed. Charges begin on November 1, 2026 at 00:00 Beijing time — that is October 31 at 19:00 Moscow time.
When it's useful: the agent needs to check a competitor's price, find a fresh news item, or verify company details from public sources. When it's not needed: answers based on internal policies and the company's knowledge base — for those, a custom RAG system without internet access is a better fit.
The model has been updated under the same name: what to check
This is the most important release item for businesses. According to the footnote on the pricing page, the old names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the corresponding models have been retired: requests are handled by DeepSeek-V4.1-Flash-1017. In other words, the model in your product may have changed on October 17 without a single code change.
Why this is a risk. The new version writes more concisely — according to DeepSeek, that is one of the improvements. But if your workflow expects a specific response format, length, or structure, something can break: a JSON parser, an email template, a required-field check.
What to check before the end of the week — editorial checklist:
- Find every place where DeepSeek is calledand list the model names being used.
- Run your test prompt set — at least 30–50 real cases with reference answers.
- Compare the format— JSON validity, required fields, response length, language.
- Compare tool calls— are the right functions being called, in the right order, with the right arguments?
- Compare the task cost— if responses got shorter, output token usage should drop; check the bill.
- Check behavior in reasoning mode and without it, if you use both.
- Log the model version for each request so the next swap is visible right away.
We explained how to organize a transition between versions without downtime in the article “Retiring LLM models: how to plan a migration”. If the model is critical to the process, keep a fallback option through a model router.
API or your own servers
Open weights under the MIT license let you deploy the model in your own environment. But this is a very heavy model. According to the Hugging Face model card on October 18:
- the weight files contain about 763 billion parameters, with a total size of about 510 GB (more precisely, 510,304,178,606 bytes; the gigabyte conversion is our editorial calculation);
- the weights are stored in mixed precision: FP8, INT8, BF16, and FP32;
- recommended generation settings:
temperature1.0,top_p1.0, context 1 million tokens,max_tokensfrom 256K; - the release does not include a chat template in Jinja format. Instead, DeepSeek provides a reference implementation in Python (
encoding.py) and a set of Rust librariesdeepseek-recipefor encoding requests; - as of October 18, the model is not available from any Hugging Face Inference Providers vendor; the community has published 33 quantized versions.
The number of parameters in the files is higher than the 552 billion mentioned in the September V4.1-Flash announcement. DeepSeek does not explain this difference; most likely, the Hugging Face counter includes all tensors in the files, but that is our assumption.
| Criterion | DeepSeek API | Your own servers or Russian cloud |
|---|---|---|
| Launch | in one day | weeks: buy or rent GPUs, set up |
| Data | goes to DeepSeek's servers | stays in your environment |
| Updates | arrive automatically, sometimes without warning | only when you decide |
| Cost | per token | for GPUs and maintenance, regardless of utilization |
| Internet search | Search and Fetch API in one account | you need a separate search service |
When a private environment makes sense and when it doesn’t, we covered in the articles “Cloud or On-Premises LLM: TCO Comparison” and “Open-Source LLM Inference in a Russian Cloud”.
Data and legal requirements
DeepSeek API is a foreign service run by the Chinese company DeepSeek. If your prompts contain personal data of customers or employees, sending that data to a foreign service falls under cross-border data transfer rules — Article 12 of Federal Law No. 152-FZ, “On Personal Data”. This is not legal advice: the decision about which data can be sent to an external API should be approved with legal counsel and the security team.
A practical approach for a pilot:
- anonymize prompts: replace full names, phone numbers, and contract numbers with tags before sending them to the API;
- separate use cases: route public data and web search through the API, and internal documents through a model in your own environment;
- check access and payment: we did not verify whether DeepSeek API is available from Russia or which payment methods work for Russian companies as of October 18.
When to choose V4.1 Flash, and when not to
The table below is the editorial methodology for making the choice, not the result of our tests.
| Situation | Choice | Why |
|---|---|---|
| An agent for code, terminal, or automation; data is not personal | V4.1-Flash-1017 via API | strong results on agent benchmarks according to DeepSeek, low cost |
| High-volume document processing without urgency | API during off-peak hours | half the price |
| The agent needs fresh internet | API + Search and Fetch | search and model in one account |
| Personal data, trade secrets | model in your own environment or a Russian model | data never leaves the perimeter |
| You need a stable, fixed version | your own servers with fixed weights | the API version may be replaced |
| Complex reasoning without tools | compare against a stronger model | on HLE, V4.1-Flash-1017 trails Opus 5 according to DeepSeek’s table |
When these recommendations don’t apply: if you don’t have a set of test tasks, build one first — without it, any model choice is still a guess. A review of Russian alternatives is in the article “LLMs in Russia 2026: GigaChat, YandexGPT, Qwen, DeepSeek”.
Frequently asked questions
What is DeepSeek-V4.1-Flash-1017?
It is an update to the DeepSeek-V4.1-Flash model, released on October 17, 2026. According to DeepSeek, it is stronger on agent, coding, and visual tasks, completes tasks faster, and produces shorter answers. The weights are open under the MIT license.
How much does DeepSeek V4.1 Flash API cost?
According to the pricing page as of October 18, 2026: during peak hours, $0.30 per 1 million input tokens without cache, $0.006 with cache, and $1.20 per 1 million output tokens. Off-peak, all prices are half as much.
When are DeepSeek peak hours in Moscow time?
From 4:00 AM to 7:00 AM and from 9:00 AM to 1:00 PM Moscow time, Monday through Friday. This is the conversion of the 1:00 AM–4:00 AM and 6:00 AM–10:00 AM UTC intervals specified by DeepSeek. Weekends are considered off-peak.
What is DeepSeek Search API?
It is an API for web search on DeepSeek’s search engine. It returns a list of pages with descriptions, dates, and relevance scores. The price is $0.75 per 1,000 successful requests, with a limit of 5 requests per second per account.
Do you need to change the code after version 1017 is released?
You do not need to change the model name: deepseek-flash and the older deepseek-v4-flash are already being served by the new version. But you should rerun your test scenarios and check response format, tool calls, and costs.
Can DeepSeek V4.1 Flash be run on your own servers?
Yes, the weights are available on Hugging Face under the MIT license. However, the weight files are about 510 GB, so you need a server cluster with a large amount of VRAM or cloud GPU rental.
Primary sources and verification method
Method: on October 18, 2026, we read the official DeepSeek pages and the model card on Hugging Face, noted the figures, dates, and terms, cross-checked them against one another, and recorded any discrepancies. Peak-hour times were converted to Moscow time from the stated UTC intervals. The cost example was prepared by the editors using hypothetical volumes. AIrassvet did not conduct its own model tests.
- DeepSeek API Docs: DeepSeek-V4.1-Flash-1017 Release, 10/17/2026 — release contents, Search and Fetch API, free-period terms.
- DeepSeek API Docs: Models & Pricing — prices, peak hours, limits, and redirects from old model names.
- DeepSeek API Docs: Search & Fetch — parameters, responses, errors, billing, and limits.
- Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash-1017 — benchmark table, generation parameters, license, weight size.
- DeepSeek API Docs: DeepSeek-V4.1-Flash Release, 09/10/2026 — V4.1-Flash family architecture.
- ConsultantPlus: Federal Law No. 152-FZ of 07/27/2006, “On Personal Data” — Article 12 on cross-border data transfers.
Limitations: the benchmarks were published by the model developer and have not been independently verified; DeepSeek may change prices and rules at any time — the company states this directly on the pricing page. Check the current terms before making a decision.
How AIrassvet helps implement DeepSeek and other LLMs
Choosing a model from a benchmark table is easy; the harder part is building it into your workflow so a version or pricing change does not disrupt operations. On the AIrassvet website , the following areas are described that directly relate to this task:
- developing AI agents integrated with communications channels, CRM systems, and accounting systems, with model version logging and response format validation;
- deploying and fine-tuning LLMs on the customer’s own servers when data must not leave to an external API;
- RAG systems built on the company’s knowledge base to answer questions from internal documents without going online;
- process audits, testing, and post-launch support, including retesting scenarios when the model is updated.
The first step is to choose one process, define the current time and quality, data sources, personal data constraints, and an acceptance criterion, such as the share of correct answers on a test set of queries.
Bottom line: a low-cost model that changes on its own
DeepSeek-V4.1-Flash-1017 starts at $0.15 per 1 million input tokens and, according to the developer, delivers a noticeable boost on agentic tasks. The Search and Fetch APIs cover a common agent need for fresh data, but their cost must be calculated separately from tokens, and the 5 requests per second limit must be built into the architecture.
The main takeaway from the release is not the benchmarks. The model under its previous name was replaced without customer involvement. Keep a test task set, log the model version, and schedule batch jobs outside Moscow peak hours — that way DeepSeek updates will deliver savings, not outages.
