DeepSeek V4.1 Flash 1017 (formal name DeepSeek-V4.1-Flash-1017) is a revised version of V4.1 Flash, which was released in September. DeepSeek published it on October 17, 2026. The API name deepseek-flash stays the same, and only the model behind it has changed. DeepSeek describes it as having "stronger benchmark performance, faster task completion, and shorter outputs than the previous version."" This article checks DeepSeek's announcement, model card, and pricing page, along with third-party evaluations and government responses, against the original sources (confirmed October 19, 2026). It separates performance, pricing, and safety. The model was released only two days ago, so there are still few third-party figures, and this article keeps DeepSeek's own numbers clearly separate from third-party numbers.

Performance

In the same range as Haiku 5.5

The third-party overall score (Artificial Analysis) is 45. This is up from 39 in the previous version, but it does not reach Claude Sonnet 5.5 (56) or Opus 5.5 (58).

Pricing

Same unit price, about half the cost per task

The unit price stays at $0.30 per million input tokens and $1.20 per million output tokens (peak hours). Output got shorter, so the cost per task fell from $0.49 to $0.26 (third-party measurement).

Safety

No evaluation of this version yet

Neither DeepSeek nor any third party has published a safety evaluation. Data is stored in China. On October 19, the United States was reported to have banned use by federal civilian agencies.

Contents

1. What changed: same name, different model

According to DeepSeek's change log, V4.1 Flash was updated to DeepSeek-V4.1-Flash-1017 on October 17, 2026. The official web and app models were switched to the same version. The API calling method and pricing stay the same. If you specify deepseek-flash, you get the new version. The Hugging Face model card calls this version a "minor update" to V4.1 Flash.

PerformanceDeepSeek announced large gains over the previous version in reasoning, coding, agent tasks, and image evaluations (figures are in Section 3).
Speed and output lengthDeepSeek says it reduced the time and output length needed for tasks, so tasks finish faster and cost less. No announcement gives a figure for how much it was reduced.
API namedeepseek-flash stays the same, but the model behind it has changed. Programs that rely on the behavior of the previous version may produce different results even without any configuration change. The old name deepseek-v4-flash is still accepted, but it now runs this version and is billed at the Flash rate.
Launched the same dayDeepSeek made its Search API, which uses its own search engine, and its Fetch API, which retrieves web pages, generally available (pricing is in Section 5).

Sources: Change Log, DeepSeek-V4.1-Flash-1017, and Models & Pricing (DeepSeek API Docs); DeepSeek-V4.1-Flash-1017 (Hugging Face). Confirmed October 19, 2026.

2. Specifications: context, thinking, size

ItemDetails
Context length and maximum output1M tokens and 384K tokens.
InputText and images.
Thinking modeOn by default. Strength has three levels: low, high, and max. The default is high.
Supported API formatsOpenAI format, Anthropic format, and Responses API. Supports JSON output and tool calling.
Parameter countThe official figures for the September V4.1 Flash are 552B for the full model. When processing input, 8B parameters are active per token. When generating output, 16B are active. There is also a separate 196B memory system (Engram).
WeightsReleased under the MIT license. The total size of the files is about 510GB.

Sources: Models & Pricing and Thinking Mode (DeepSeek API Docs); DeepSeek-V4.1-Flash and DeepSeek-V4.1-Flash-1017 (Hugging Face). Confirmed October 19, 2026. The model card for 1017 does not state a parameter count. Hugging Face's automatic count from the weight files is about 763B. DeepSeek has not published the reason for the difference.

There are two rules that often cause problems when using the API. First, in conversations that use tools, you must return the model's reasoning content (reasoning_content) in full with the next request. If you don't, you get a 400 error. Second, in thinking mode, settings such as temperature have no effect. They are ignored without an error. If you specify medium or xhigh as the thinking strength, each is treated as high.

Source: Thinking Mode (DeepSeek API Docs, confirmed October 19, 2026).

3. Performance (1) Numbers published by DeepSeek

The comparison table from the model card is reproduced below. All of these numbers were measured by DeepSeek itself.

Evaluation1017Previous version (September)Kimi K3GLM-5.3GPT-5.6 SolClaude Opus 5
HLE (no tools)42.636.843.542.0*44.556.3
Terminal-Bench 4.042.331.212.637.939.951.8
DeepSWE v1.177.074.267.566.973.074.0
ProgramBench29.120.317.519.023.037.0
NL2Repo-Bench68.265.458.058.056.875.3
Automation-Bench v1.0.660.554.846.748.845.850.3
Agents' Last Exam35.431.827.628.526.728.6
Apex Bench v1.172.164.647.856.654.365.8
Chartography (with tools)83.078.968.1—79.984.0
BabyVision (with tools)92.589.685.7—88.994.1
ZeroBench-main (with tools)51.049.041.0—53.052.0

Sources: DeepSeek-V4.1-Flash-1017 (Hugging Face model card) and Change Log (confirmed October 19, 2026). * marks the score for the text-only part of HLE. The model card notes that Terminal-Bench 4.0 was measured with Claude Code and DeepSWE v1.1 with the Mini-SWE-V2 execution framework.

There are four things to keep in mind when reading this table.

  1. The comparison models are a generation older. The table includes Claude Opus 5 and GPT-5.6 Sol, but not the current Claude Opus 5.5 or GPT-6 models. Compared with Opus 5, the model beats it on DeepSWE, but it falls short on HLE, Terminal-Bench 4.0, ProgramBench, and NL2Repo-Bench.
  2. The conditions for a fair comparison are not stated. The thinking strength settings (low, high, max) are not given. It is also not stated whether the previous version was measured under the same conditions.
  3. Some figures from the previous version are no longer reported. The September version also reported GPQA Diamond, Codeforces, math, and cybersecurity evaluations, but 1017 does not report them.
  4. DeepSeek's own figures tend to come out high. The US standards body NIST's CAISI evaluated V4 Pro in April 2026. It reported that DeepSeek's own evaluation gave higher scores than CAISI's (see Section 6). Until independent measurements for 1017 are available, the safest approach is to read these figures with caution.

4. Performance (2) Third-party evaluations

The third-party overall score available so far is the Intelligence Index from Artificial Analysis. It combines the results of several tests into one score. Each thinking strength setting appears as a separate row (the setting is shown in parentheses). The values below are from the Artificial Analysis listing on October 19, 2026, shown next to nearby models.

Sources: Artificial Analysis model leaderboard and DeepSeek V4.1 Flash 1017 page (confirmed October 19, 2026). The model was released only recently, and some pages on the same site still show values for the previous version. Values may change over time.

Three things can be said from this.

  • The score rose 6 points from the previous version, and the cost per task fell to almost half. On the Artificial Analysis page, running the full evaluation used 210M output tokens. That is more than the median of the compared models (140M), and the site describes it as "very verbose."
  • Both the score and the cost are close to Claude Haiku 5.5. Compared with GLM-5.3, which scores the same 45, and Kimi K3, which scores 44, the cost per task is about one-eighth.
  • The gap with Sonnet 5.5 and Opus 5.5 is more than 10 points. The cost is about one-twentieth, so the choice depends on whether you want to assign difficult work or run high-volume work cheaply.

On the DeepSeek API, speed was about 225 tokens per second, and the first token arrived in about 1.1 seconds (same page). That is nearly the same speed as the previous version, so the claim that tasks "finish faster" is likely due to shorter output. As of October 19, we could not find results for 1017 on the official SWE-bench or Terminal-Bench leaderboards.

5. Pricing: peak rates apply during Japan's daytime

deepseek-flash (per million tokens, US dollars)Off-peakPeak
Input (cache hit)$0.003$0.006
Input (cache miss)$0.15$0.30
Output$0.60$1.20

Source: Models & Pricing (DeepSeek API Docs, confirmed October 19, 2026). The update to 1017 did not change pricing (Change Log).

Peak hours match Japan's working hoursPeak hours are 1:00 to 4:00 and 6:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. In Japan Standard Time, these are weekday hours of 10:00 to 13:00 and 15:00 to 19:00. Evenings and weekends are half price.
The cache makes a large differenceInput that hits the cache costs one-fiftieth of input that misses it. For agent use that sends the same preamble every time, whether the cache is used has a large effect on cost. Caching is automatic, but a hit is not guaranteed.
Search and fetch are billed separatelyThe Search API and Fetch API cost $0.75 per 1,000 successful calls. Calls made on the server side from Claude Code, Codex, and DeepSeek Harness are free for now. Billing starts at 16:00 UTC on October 31, 2026, which is 1:00 on November 1 in Japan Standard Time.
Granted balance is used firstIf you have both a free granted balance and a deposited balance, the granted balance is used first.

Sources: Models & Pricing, Search & Fetch, Context Caching, and DeepSeek-V4.1-Flash-1017 (DeepSeek API Docs, confirmed October 19, 2026).

You can also use the model outside DeepSeek's official API. Relay services such as OpenRouter offer 1017. Some charge the same unit price as the official API, and some charge less. However, many of the cheaper routes run a compressed (quantized) version of the weights, so the quality and speed of answers may differ from the official version. How data is handled depends on each company's terms.

6. Safety: no evaluation yet, what is known

As of October 19, 2026, neither DeepSeek nor any third party has published an evaluation of 1017 that measures how it responds to harmful requests, its resistance to jailbreaks, or how hard it is to hijack as an agent. The announcement, the model card, and the September technical report do not include a safety evaluation section. What is known comes from the previous generation and from the company as a whole. These findings do not necessarily apply to 1017.

Evaluation of V4 Pro (May 2026)NIST's CAISI reported that V4 Pro's capabilities trail the US frontier by about eight months, and that DeepSeek's own evaluation gave higher scores than CAISI's. Safety (jailbreaks and hijacking) was not covered by this evaluation.
Evaluation of the previous generation (September 2025)The same CAISI report said that, for R1-0528, when well-known jailbreak techniques were used, the model complied with 94% of clearly malicious requests (the US comparison model complied with 8%). When used as an agent, it followed hijacking instructions at a rate 12 times higher on average than leading US models. It also produced inaccurate content aligned with Chinese Communist Party claims at four times the rate of US models.
Versions with safety features removed already existBecause the weights are public, third parties have reportedly released versions of V4.1 Flash with the refusal behavior removed. If you run the model locally, you cannot rely on the model's refusals for safety.

Sources: CAISI Evaluation of DeepSeek V4 Pro (NIST, May 2026); CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks (NIST, September 30, 2025); AlphaSignal (report on the version with safety features removed; the figures are self-reported by the publisher). Confirmed October 19, 2026.

When you use the model from an agent such as Claude Code, the most effective step is not to wait for a safety evaluation of the model. Instead, limit what the agent is allowed to do. Three basics are: don't use settings that approve actions automatically, don't let the agent read files that contain secrets, and don't trust instructions written in content fetched from the web.

7. Where data goes and how countries are responding

When you use DeepSeek's official API or app, the content you send reaches DeepSeek's servers. DeepSeek's privacy policy (Japanese version, last updated February 10, 2026) says it stores the information it collects on "secure servers located in the People's Republic of China." One of the purposes it lists is "training and improving our technology, such as machine learning models and algorithms." The developer terms (effective April 29, 2026) leave the handling of personal information to this policy and specify the laws of mainland China as the governing law. Neither document includes a promise that content sent through the API will not be used for training.

Sources: DeepSeek Privacy Policy and DeepSeek Open Platform Terms of Service (confirmed October 19, 2026. When the privacy policy is opened from Japan, the Japanese version is shown).

Country or agencyResponse
Japan (Personal Information Protection Commission)On February 3, 2025, the Commission issued "Information on DeepSeek." It warned that data is stored on servers in the People's Republic of China and that Chinese laws (including the Personal Information Protection Law, the Cybersecurity Law, the Data Security Law, and the National Intelligence Law) apply. It does not prohibit private use.
United States (Department of Homeland Security)On October 19, 2026, the department was reported to have issued an order to remove DeepSeek's products, and third-party services that use DeepSeek, from the IT systems of federal civilian agencies, and to ban their procurement. The order is scheduled to be published in the Federal Register on October 20. At the time this article was written, we had not verified the full text of the order or its scope.

Sources: Information on DeepSeek (Personal Information Protection Commission, February 3, 2025, updated March 5, 2025); US Moves to Bar DeepSeek From Federal Agencies and Contracts (Bloomberg, October 19, 2026); DeepSeek Blocked In US Federal Agencies (BankInfoSecurity); Federal Register public inspection listing.

8. Choosing how to use it: API, Claude Code, running locally

DeepSeek's official APIThe cheapest option, and the new version is used automatically. Content you send is stored on servers in China. It suits work that is fine to share externally, such as open-source projects or personal study.
Using it from Claude CodeDeepSeek officially provides setup instructions for using the model with Claude Code. We cover the setup steps and cost considerations in Using DeepSeek, GLM, and Kimi with Claude Code. Note that search billing starts on November 1.
Using a relay serviceServices such as OpenRouter provide access. How data is handled depends on that company's terms. Cheaper routes often use quantized versions.
Running the weights locallyYour code never leaves your machine. However, the weight files alone are about 510GB, and running them requires an environment with several data center GPUs. This is not something you can run on a typical PC. For models you can run locally, see our comparison of recommended local LLM models.

Summary

  • DeepSeek V4.1 Flash 1017 is a revised version of V4.1 Flash, released on October 17, 2026. The API name deepseek-flash stays the same while the model behind it has changed, and pricing is unchanged.
  • DeepSeek's announcement says the model beats the previous version on all 11 evaluations. However, the comparison models are a generation older, and the test conditions are not described in detail.
  • Artificial Analysis, a third-party source, gives it 45 points (the previous version had 39). That places it in roughly the same range as Claude Haiku 5.5 (43 points) and well below Sonnet 5.5 and Opus 5.5, which are more than 10 points higher.
  • The cost per task fell from $0.49 to $0.26. The unit price is $0.30 per million input tokens and $1.20 per million output tokens (peak hours), and weekday daytime in Japan is peak time.
  • Search and fetch calls made from Claude Code and similar tools cost $0.75 per 1,000 successful calls, starting at 1:00 on November 1 Japan Standard Time.
  • No safety evaluation of 1017 exists yet. The previous generation was rated as vulnerable to jailbreaks and hijacking. Data is stored on servers in China, and the United States has reportedly banned its use by federal civilian agencies.

FAQ

Q. Can I keep using the previous V4.1 Flash?

A. Not through DeepSeek's official API. Both deepseek-flash and the old name deepseek-v4-flash now run 1017. If you want the previous version, you would need to run its publicly available weights yourself, or find a relay service that still offers it.

Q. Did the price go up?

A. The unit price has not changed. Output is shorter, so the cost of the same task goes down. In Artificial Analysis's measurements, the average cost per task fell from $0.49 to $0.26. However, starting November 1 (Japan Standard Time), search and fetch calls made from Claude Code and similar tools will incur a separate charge.

Q. Is it safe to send my work code?

A. Content sent to the official API is stored on servers in China, and the terms do not include a promise that it won't be used for training. Work code, customer data, and secrets should not be sent unless your employer has approved it.

Q. Is it banned in Japan?

A. As of October 19, 2026, we found no Japanese measure that prohibits private use. The Personal Information Protection Commission has warned that data is stored on servers in China and that Chinese laws apply.

Sources