🔍 Read the full analysis: Can DeepSeek V4.1 Flash 1017 Deliver Near-Frontier Agents For Less? on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
DeepSeek released V4.1 Flash 1017 on October 17, replacing the September model under the existing API name and price. DeepSeek reports improvements on all 11 of its published benchmarks, with larger gains on some agent tasks; an early seven-task independent pilot supports the cost case but is too small to establish a broad ranking. Independent benchmark reproduction and direct comparisons with newer frontier models remain outstanding.
The new version is served under deepseek-flash; DeepSeek says the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp route to it as well. The model’s weights are available under the MIT license. DeepSeek describes the release as a minor update and does not detail what changed in training.
In DeepSeek’s model-card results, Terminal-Bench 4.0 rose from 31.2 to 42.3, while ProgramBench increased from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark, moved from 74.2 to 77.0. These are company-reported scores; benchmark results depend on evaluation setup, and the listed tests used different harnesses, including Claude Code for Terminal-Bench and Mini-SWE-V2 for DeepSWE.
DeepSeek also says the model produces shorter outputs, which can lower the time and token cost of a task. The API price remains unchanged: for example, the listed output rate is $0.60 per million tokens off-peak and $1.20 at peak. The company’s pricing page defines weekday peak periods as 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays.
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 Agent Work
The release’s practical case is cost per completed task, not a claim that 1017 is the strongest model overall. The Decoder reported DeepSeek’s DeepSWE cost chart puts 1017 at 77% performance for just under $0.20 per task; a lower-effort setting scored above 72% at about $0.10. On that chart, the September version cost about $0.50 per task at 74.2%, while the compared frontier models ranged from about $2 to more than $10. These are reported figures tied to that benchmark and setup, not a guarantee of costs on other workloads.
For developers running many coding or tool-use requests, even a modest drop in cost per successful task can matter more than a leaderboard position. DeepSeek’s own table gives 1017 the lead in four of 11 rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. That makes the result relevant as a lower-cost option to test, rather than evidence that it broadly matches or surpasses the current frontier.
Kingy AI’s early pilot offers a limited outside check on the efficiency argument. Across seven small tasks run twice, 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, with 1017 reportedly costing about a quarter as much. The pilot found differences in speed, including one 1017 task that took about 27 seconds versus roughly three seconds for Sonnet. With all three models scoring perfectly on a small sample, the test cannot establish which performs better overall.
AI model performance benchmarking tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the October Update Fits
DeepSeek released the earlier V4.1 Flash on September 10. The October 17 release replaces that version for users of the company’s hosted services; callers using the standard API model name receive 1017 automatically, according to the source report. The open weights give developers another route to test or deploy the model, subject to their own hardware and evaluation needs.
DeepSeek’s technical report describes a sparse Mixture of Experts model with image input and a one-million-token context window. It gives a 552-billion-parameter backbone, with about eight billion active parameters during prefill and 16 billion during decoding. Hugging Face lists a different total, 763 billion, reflecting a different counting method. The hosted API is reported to support up to 384,000 output tokens.
The report also describes a Causal Encoder-Decoder design and KV-cache compression. The Decoder attributed the benchmark gains to post-training with reinforcement learning and model self-feedback, but DeepSeek has not publicly specified the training changes behind 1017. The company says it tuned the release for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy.
““minor update””
— DeepSeek, in its release description
As an affiliate, we earn on qualifying purchases.
Independent Tests Still Needed
DeepSeek produced the 11 benchmark scores in its comparison table, and the source report says they have not been independently reproduced. The company’s chosen rivals include GPT-5.6 Sol and Claude Opus 5, but not newer models named in the source, such as GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon. Separately reported Terminal-Bench results for current frontier models were said to fall between 57 and 66, but those evaluations used different setups and are not directly comparable.
The meaning of the reported gains outside DeepSeek’s tests is also uncertain. The company has not explained the training changes, and shorter outputs do not guarantee lower total cost if a task needs retries, additional tools or more input tokens. The independent pilot covered only seven tasks, run twice; its perfect pass counts across three models do not settle reliability on larger or more varied workloads. More detail on the additional search-tool announcement referenced in the source is not available in the material provided.
As an affiliate, we earn on qualifying purchases.
Watch for Broader Evaluations
The next useful evidence will be independent, repeatable evaluations that disclose prompts, tools, model settings, task selection, failure rates, latency and total cost. Developers can also compare 1017 with current alternatives on their own workloads, since tool setup and the number of retries can change the cost of a completed task. DeepSeek recommends version 0.2.1-rc.1 or later of its harness for its own agent setup.
For API users, the model-name change means the new version is already the default under deepseek-flash, so teams relying on the older model may want to check their evaluations and output behavior. The available source does not give a date for a follow-up model release or a schedule for additional benchmark results. Until those arrive, 1017’s strongest supported case is a company-reported improvement on selected agent tests at unchanged listed prices, backed by a small outside pilot rather than broad independent validation.
As an affiliate, we earn on qualifying purchases.
Key Questions
What changed with DeepSeek V4.1 Flash 1017?
Released on October 17, 2026, 1017 replaced the September V4.1 Flash on DeepSeek’s website, app and API. DeepSeek reports higher scores on all 11 benchmarks in its model card and says outputs are shorter. The API price was unchanged.
Is 1017 proven to outperform frontier models?
No. The benchmark figures are reported by DeepSeek and have not been independently reproduced, according to the source report. In DeepSeek’s table, 1017 leads four of 11 categories; some newer frontier models are not included.
How much can an agent task cost with 1017?
The Decoder reported that DeepSeek’s DeepSWE chart showed a cost of just under $0.20 per task at a 77% score, and about $0.10 at a lower-effort setting scoring above 72%. Those figures apply to that benchmark and setup; actual costs vary with tokens, tools and retries.
Do API users need to change the model name?
DeepSeek says requests using deepseek-flash now route to 1017 automatically. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are also routed to the update, according to the source report.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
