🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: A Table-First Guide To Switching on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
DeepSeek released V4.1 Flash 1017 on Oct. 17, replacing the September model on its site, app and API without changing prices. Its published results show gains across 11 benchmarks, with larger improvements on some agent tasks, but the figures are company-reported and its comparisons use selected, older rivals. The update’s case rests on lower cost per completed task, while independent evidence remains limited.
The update is served under the API name deepseek-flash; DeepSeek says older names, including deepseek-v4-flash and deepseek-v4-flash-vision-exp, also route to the new version. The model’s weights are available under the MIT license. DeepSeek describes 1017 as a minor update and has not detailed what changed in its training.
In DeepSeek’s model-card results, Terminal-Bench 4.0 increased from 31.2 for the September model to 42.3, while ProgramBench rose from 20.3 to 29.1. The gain on DeepSWE v1.1 was smaller, from 74.2 to 77.0. The company says outputs are shorter, which can reduce completion time and token costs; it has not published independent measurements confirming those claims.
The listed API rates vary by time and cache status. At off-peak rates, input costs $0.15 per million tokens when uncached and output costs $0.60 per million. Peak rates are $0.30 and $1.20, respectively; cached input is cheaper. DeepSeek defines peak weekday periods as 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. These rates make the cost of a particular job dependent on its token use, cache hits and timing.
Infographic · DeepSeek V4.1 Flash 1017
Better on every benchmark, a fraction of the cost per task, and still behind the frontier.
DeepSeek’s October update improves all eleven scores it published over the September version. Against the four rivals DeepSeek picked, it leads four benchmarks and trails seven. Its real advantage is the price of a finished task.
1017 vs the September version
Score per benchmark (0–100), sorted by gain. Hover a row for the exact values.
- Terminal-Bench 4.042.331.2+11.1
- ProgramBench29.120.3+8.8
- Apex Bench v1.172.164.6+7.5
- HLE (no tools)42.636.8+5.8
- Automation-Bench60.554.8+5.7
- Chartography83.078.9+4.1
- Agents' Last Exam35.431.8+3.6
- BabyVision92.589.6+2.9
- DeepSWE v1.177.074.2+2.8
- NL2Repo-Bench68.265.4+2.8
- ZeroBench-main51.049.0+2.0
Against the best rival per benchmark
Points ahead of or behind the strongest of Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5 on each row.
- Automation-Benchbest rival: Claude Opus 5 50.3+10.2
- Agents' Last Exambest rival: Claude Opus 5 28.6+6.8
- Apex Bench v1.1best rival: Claude Opus 5 65.8+6.3
- DeepSWE v1.1best rival: Claude Opus 5 74.0+3.0
- Chartographybest rival: Claude Opus 5 84.0-1.0
- BabyVisionbest rival: Claude Opus 5 94.1-1.6
- ZeroBench-mainbest rival: GPT-5.6 Sol 53.0-2.0
- NL2Repo-Benchbest rival: Claude Opus 5 75.3-7.1
- ProgramBenchbest rival: Claude Opus 5 37.0-7.9
- Terminal-Bench 4.0best rival: Claude Opus 5 51.8-9.5
- HLE (no tools)best rival: Claude Opus 5 56.3-13.7
Cost per DeepSWE task
USD per task, linear scale to $10, as read from DeepSeek’s chart.
- 1017, low effortabove 72% of DeepSWE solvedabout $0.10
- 1017, default77.0% of DeepSWE solvedjust under $0.20
- V4.1 Flash (September)74.2% of DeepSWE solvedabout $0.50
- Frontier models on the chartGPT-6 Astra, Gemini 3.8 Flash, Claude Opus 5$2 to $10+
None of the frontier models on DeepSeek’s chart reached 77%.
API price per 1M tokens
| Token type | Off-peak | Peak |
|---|---|---|
| Input, cache miss | $0.15 | $0.30 |
| Input, cache hit | $0.003 | $0.006 |
| Output | $0.60 | $1.20 |
Peak: weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays.
All scores are DeepSeek’s own and not independently reproduced; rivals are the ones DeepSeek chose (GPT-5.6 Sol and Claude Opus 5, not the current GPT-6 and Opus 5.5 models). Sources: DeepSeek model card, DeepSeek API pricing, cost per task via The Decoder (17–18 Oct 2026).
Powered by Thorsten Meyer AI
Lower Costs for Long Agent Runs
The update is aimed at developers who run coding and other agent tasks repeatedly, where cost per completed task can matter more than a model’s place on a general leaderboard. DeepSeek’s published DeepSWE comparison, as reported by The Decoder, puts 1017 at 77% for just under $0.20 per task; a lower-effort setting reportedly scores above 72% at about $0.10. These are figures reported from DeepSeek’s evaluation, not a broad independent cost study.
That distinction matters because lower token prices do not by themselves prove that a model completes work more cheaply or reliably. A task may require retries, more tokens or human correction. A small pilot by Kingy AI offers an early, limited check: across seven tasks run twice, 1017 passed all 14 runs, as did GPT-6.1 Sol and Claude Sonnet 5.5, at roughly a quarter of their cost, according to Kingy AI. The test was too small to rank the models, and one task took 1017 about 27 seconds compared with roughly three seconds for Sonnet.
DeepSeek’s own table also shows limits to the model’s standing. It leads four of 11 rows; Claude Opus 5 leads six, and GPT-5.6 Sol leads one. The update may be a cost-effective option for some workloads, but the available evidence does not establish that it outperforms newer competitors overall.
As an affiliate, we earn on qualifying purchases.
What the Benchmark Table Covers
DeepSeek released the previous V4.1 Flash version on September 10, 2026. The October 17 version takes its place as the default on DeepSeek’s services, so API users calling the listed model name receive 1017 without changing the name. The company recommends its DeepSeek Harness version 0.2.1-rc.1 or later and says it tuned the model for that harness and Tencent’s CodeBuddy and WorkBuddy.
The model card describes a sparse Mixture of Experts model with image input and a one-million-token context window. DeepSeek’s technical report, as summarized by Kingy AI, describes a 40-layer Causal Encoder-Decoder architecture with 384 routed experts, six active per token, plus one shared expert. Sources differ in how they count model size: the report gives a 552-billion-parameter backbone, while Hugging Face lists 763 billion. The hosted API is listed as allowing up to 384,000 output tokens.
Benchmark scores require care. DeepSeek’s table includes its September model and selected rivals, not every current model. Its Humanity’s Last Exam result uses the full set, while the text-only results are different; the table’s GLM-5.3 figure is text-only. Terminal-Bench was run in the Claude Code harness and DeepSWE in Mini-SWE-V2. Separately reported results for newer frontier models are not direct comparisons because evaluation conditions differ.
““minor update””
— DeepSeek
As an affiliate, we earn on qualifying purchases.
Independent Checks Are Still Limited
Independent replication of DeepSeek’s benchmark results is not available in the source material. The comparisons also leave open how 1017 performs against newer models under matched test conditions. DeepSeek has not explained what training changes produced the gains, and its claim that shorter outputs lower costs needs to be considered alongside retries, latency and the amount of human correction a task requires.
The Kingy AI pilot was small, covering seven tasks with two runs each. Its perfect scores across three models cannot establish a general performance ranking, and one timing example showed 1017 taking longer than Sonnet. The full terms and wider applicability of the pilot are not established here. It is also unclear how often the reported savings hold across varied workloads, prompt lengths, cache rates and peak-hour pricing.
As an affiliate, we earn on qualifying purchases.
Watch for Reproductions and Usage Data
Developers can access 1017 through DeepSeek’s existing website, app and API routes, or download the MIT-licensed weights from Hugging Face and ModelScope. The next useful evidence will be independent benchmark runs that use the same harness, task set and model settings across current competitors, alongside published cost and latency measurements.
For teams considering a switch, the practical next step is to test their own representative tasks and account for token use, cache hits, retries and the UTC pricing schedule. DeepSeek has not announced a further release date or a detailed training disclosure in the available material. Until those details or broader evaluations appear, the company’s results describe its own testing rather than a settled comparison across the market.
As an affiliate, we earn on qualifying purchases.
Key Questions
What changed in DeepSeek V4.1 Flash 1017?
DeepSeek says the October 17 update improves all 11 benchmarks in its model card compared with the September version. The larger listed gains include Terminal-Bench 4.0 and ProgramBench. The company has not provided a detailed account of the training changes.
Do API users need to change the model name?
DeepSeek says calls to deepseek-flash now receive 1017 automatically. It also says the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp route to the updated model.
Is 1017 proven to outperform competing models?
No. The published benchmark results are DeepSeek-reported and have not been independently reproduced in the source material. Its table uses selected competitors, and some newer-model results cited elsewhere were measured under different conditions.
How much does using the model cost?
DeepSeek lists off-peak prices of $0.15 per million uncached input tokens and $0.60 per million output tokens. Peak rates are higher, while cached input costs less. A job’s actual cost depends on token volume, cache status and timing.
What should teams test before switching?
Teams should compare 1017 with their current model on representative tasks, tracking completion rate, latency, retries and total cost, not just token prices or benchmark scores. The small Kingy AI pilot is not enough to predict results for a broader workload.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
