InstantVid
All AI news
Model updates

DeepSeek V4.1 Flash 1017: Benchmarks, Pricing & Setup

Compare DeepSeek V4.1 Flash 1017 with the previous model: benchmark gains, unchanged API prices, agent improvements, official access and a practical API quickstart.

InstantVid · EditorialPublished

DeepSeek-V4.1-Flash-1017 was released on October 17, 2026. This minor update improves reported reasoning, coding and agent results while reducing task runtime and output length. The official website, app and API have been updated. The API name remains deepseek-flash, with no change to the calling method or model rates. See the release announcement and change log.

This model snapshot and comparison was checked on October 18, 2026. Benchmark numbers are publisher-reported. Score differences and cost examples are our calculations, not independent model-test results.

DeepSeek 4.1 Flash 1017 at a glance

Item What to know
Release name DeepSeek-V4.1-Flash-1017
Release date October 17, 2026; a minor update to V4.1-Flash
Main workflows Reasoning, coding and agents, with visual understanding
Official API ID deepseek-flash, currently serving 1017
Input and output Text/image understanding and text generation; not video generation
Current API limits 1M context and up to 384K output, subject to the total context limit
Access Official web, app, API and open weights

Check the current model table for API limits and the preceding model card for the model family's input/output description. A large context window is a capacity limit, not a promise that filling it will improve every task.

Benchmark results: Flash-1017 vs the previous V4.1-Flash

These figures consistently use the version comparison in the 1017 model card. The delta is the new score minus the old score, rather than a percentage improvement in overall intelligence or speed.

Benchmark V4.1-Flash Flash-1017 Score change
HLE, no tools 36.8 42.6 +5.8
HLE, text-only subset, no tools 39.1 44.0 +4.9
Terminal-Bench 4.0 31.2 42.3 +11.1
DeepSWE v1.1 74.2 77.0 +2.8
ProgramBench 20.3 29.1 +8.8
NL2Repo-Bench 65.4 68.2 +2.8
Automation-Bench v1.0.6 54.8 60.5 +5.7
Agents’ Last Exam 31.8 35.4 +3.6
Apex Bench v1.1 64.6 72.1 +7.5
Chartography, with tools 78.9 83.0 +4.1
BabyVision, with tools 89.6 92.5 +2.9
ZeroBench-main, with tools 49.0 51.0 +2.0

There are 11 main metrics, with the HLE text-only subset displayed separately. Use 42.6 for the main HLE entry; 44.0 describes a narrower subset. ZeroBench-main uses Pass@5 and should not be averaged with Pass@1 results. The release notes identify Claude Code for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE.

One source discrepancy matters: the older model card reports 64.0 for NL2Repo, while the new comparison and change log use 65.4 as the earlier score. We use the latter consistently, giving a 2.8-point increase rather than combining sources to claim 4.2. The public documents do not fully reconcile that difference.

What improved in reasoning, coding and agents?

Reasoning: keep the evaluation scope visible

The HLE increase from 36.8 to 42.6 supports a stronger result on that reported evaluation. It does not establish equal gains in every subject or guarantee that a plausible explanation is correct. In your own trial, ask for assumptions, missing information and a way to check the final answer.

Coding: prioritize terminal tasks and completed projects

Terminal-Bench 4.0 increases by 11.1 points and ProgramBench by 8.8. Calculated relative to their earlier scores, those gains are about 35.6% and 43.3%. They are relative changes in benchmark scores, not evidence that programming is 43.3% faster.

A useful practical test is to fix a real bug or implement a small feature and pass its tests. Review the changed files, unnecessary dependencies and edge cases. A polished first response is less informative than a working, appropriately scoped result.

Agent workflows: measure completion rather than tool activity

The Automation-Bench, Agents’ Last Exam and Apex Bench improvements make multi-step workflows worth retesting. Our recommendation is to record task completion, human corrections and failure recovery. More tool calls alone are not evidence of a better agent.

Training and efficiency: distinguish this update from the older architecture

The launch announcement attributes stronger end-to-end project development to reinforcement learning with self-feedback, and demonstrates game and 3D-scene projects. Those are publisher demonstrations, not a measured success rate for your repository.

DeepSeek also reports shorter outputs and faster task completion, without publishing one reduction percentage that applies to every workload. Do not repackage all of the V4.1 family's earlier architecture and caching features as changes introduced on October 17.

How does Flash-1017 compare with other models?

The following selection retains the rival versions in DeepSeek's release comparison. It is not our independent leaderboard and does not establish results against newer versions that were not in that table.

Benchmark Flash-1017 GPT-5.6 Sol Opus 5
HLE, no tools 42.6 44.5 56.3
Terminal-Bench 4.0 42.3 39.9 51.8
DeepSWE v1.1 77.0 73.0 74.0
ProgramBench 29.1 23.0 37.0
Automation-Bench v1.0.6 60.5 45.8 50.3
Agents’ Last Exam 35.4 26.7 28.6
Apex Bench v1.1 72.1 54.3 65.8

In the complete publisher table, 1017 scores above the other listed models on four of 11 main metrics. It does not lead HLE, Terminal-Bench 4.0 or ProgramBench. A defensible reading is that it is competitive on selected agent and development tasks, rather than universally superior.

Harnesses matter. The framework notes attached to the new release are not identical to every historical configuration in the older card. Model, prompt, tools, reasoning budget and stopping rules all affect results; a release table does not replace a controlled A/B evaluation in your own workflow.

Pricing changes: unchanged token rates, potentially lower task costs

The October 17 change log explicitly says API pricing is unchanged. The current USD rates below were checked on October 18 and are per one million tokens, separate from consumer subscriptions.

Usage Off-peak Peak Change in this update
Input, cache hit $0.003 $0.006 None
Input, cache miss $0.15 $0.30 None
Output $0.60 $1.20 None

The official pricing page defines peak hours as Monday–Friday 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. All other hours, including weekends, are off-peak. These windows are 09:00–12:00 and 14:00–18:00 Beijing time.

For a hypothetical 100,000 uncached input tokens and 10,000 billed output tokens, the listed items cost 0.1 × $0.15 + 0.01 × $0.60 = $0.021 off-peak, or $0.042 at peak. This is arithmetic for specified usage, not an observed run or fixed-price task.

Fewer output tokens for the same accepted result can reduce the bill. Retries, additional tool use and incomplete work can increase it. Compare the total cost of an accepted task, including billed reasoning, caching, tools and failed attempts.

Search and Fetch are separate services with their own charges

The Search and Fetch guide distinguishes finding web pages from retrieving a page's content. These are separate APIs; an ordinary chat request does not automatically become a web-search workflow. Each endpoint's standard USD rate is $0.75 per 1,000 successful requests. The Chinese announcement lists RMB 5 per thousand; check your account's settlement currency rather than treating this article as a currency conversion.

Server-side Search/Fetch calls made while using the DeepSeek API through DeepSeek Harness, Claude Code or Codex are temporarily unbilled. That arrangement ends at 00:00 Beijing time on November 1, 2026, or 16:00 UTC on October 31. It does not make model tokens free or cover every standalone Search/Fetch request. See the billing notice.

How to start using DeepSeek V4.1 Flash 1017

Web and app users

Open DeepSeek Chat, or use the official downloads page for the mobile app. DeepSeek says these services have been updated, although a particular interface need not display the complete 1017 suffix.

Try a narrowly scoped task: “Turn these product facts into a 15-second ad script. First identify unsupported claims, then suggest three shots.” Check factual and product-benefit claims yourself; fluency is not verification.

Developers: keep the API ID and test a small request

  1. Sign in to the DeepSeek API platform, review billing and create a key.
  2. Set DEEPSEEK_API_KEY in a trusted terminal. Keep it out of public pages and repositories.
  3. Follow the official quickstart using deepseek-flash. Do not assume the full release name is a supported official API model value.
  4. Check the reply, finish reason and usage before extending the workflow.

This original connectivity example becomes billable when you run it with a real key. We checked the request structure without making a paid API call.

curl --fail-with-body --max-time 120 https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "deepseek-flash",
    "messages": [
      {"role": "user", "content": "Write three product-demo lines for a stainless-steel cup. Do not invent insulation duration or certifications."}
    ],
    "thinking": {"type": "enabled"},
    "reasoning_effort": "high",
    "max_tokens": 4096,
    "stream": false
  }'

The 4,096-token setting is a small example budget, not a benchmark-reproduction configuration. If finish_reason is length, do not accept a truncated response as completed work. Review the budget, task and API parameter reference. Consider max reasoning effort only after evaluating its tradeoffs.

The temporarily compatible legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to 1017. They cannot pin the previous model for an A/B comparison. For historical reproduction, verify that a provider actually retains the intended checkpoint and record its configuration.

Agent users: check the Harness version and workspace

The announcement recommends DeepSeek Harness 0.2.1-rc.1 or newer. Get the client and developer links from the official Harness page. Web UI users can follow its quickstart: configure an API key, select a workspace and begin a session. Start with a small, verifiable task in a project copy, then inspect changes and tests.

Open-weight users

DeepSeek links to Hugging Face and ModelScope. The model card identifies the MIT license and provides encoding and inference instructions. Before self-hosting, check hardware capacity, inference support and prompt formatting. A download link is not evidence of a tested one-click installation on a consumer computer.

Our brief assessment: a useful update to verify in your own workflow

Teams already using deepseek-flash should prioritize regression checks on terminal coding, project development and multi-step agent tasks. Web users can try a small task directly. Developers should establish whether existing workflows still finish correctly after the automatic version change.

Our documentation-based assessment is positive about the breadth of reported gains, unchanged rates and clear access path. The release does not establish your own success rate, tail latency or savings per accepted task. Use five fixed tasks and record acceptance, elapsed time, all billed tokens, tool charges and human corrections before making a migration decision.

Video creators can use a language model to prepare scripts, shot lists and source material, then continue in our video studio or product advertising workspace. This article does not announce a DeepSeek video-model integration. Actual generation models and credit quotes are shown in the studio and on the pricing page.

The cover is an original AI-assisted editorial illustration, not an official product screenshot. Follow the primary sources for later changes to aliases, rates and availability.

Sources & further reading

Sources reviewed Oct 18, 2026