🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Compare Agent Performance Before You Switch 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 across its site, app and API at the same listed token prices. Its own benchmarks show gains on all 11 reported tests, particularly some long-running agent tasks, but the results have not been independently reproduced and its comparison set omits newer rivals.
The update is served under the API name deepseek-flash; the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to it, according to the source report. DeepSeek describes 1017 as a minor update. Its chart compares it with the September release and four selected competitors.
In DeepSeek’s figures, Terminal-Bench 4.0 increased from 31.2 to 42.3, while ProgramBench rose from 20.3 to 29.1. The gain on DeepSWE v1.1 was smaller, moving from 74.2 to 77.0. The scores are benchmark results reported by DeepSeek, not independently verified measurements.
The source says the model’s weights are available under the MIT license on Hugging Face and ModelScope. DeepSeek lists API prices of $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, with higher peak rates. It also says output lengths are shorter, which could reduce time and cost for some tasks; actual savings will depend on usage and task results.
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 update matters most to developers running repeated coding or other agent workflows, where a model’s cost per completed task can matter more than its score on an isolated prompt. DeepSeek’s published figures suggest 1017 may improve that trade-off: its DeepSWE result is 77.0%, and the source report says DeepSeek’s cost chart places that result at just under $0.20 per task. Those figures are company-reported, so buyers should treat them as a starting point for testing rather than a guaranteed cost or performance level.
DeepSeek’s posted rates also make small-scale trials relatively inexpensive for some workloads. The source gives an example of one million uncached input tokens and 200,000 output tokens costing $0.27 off-peak at the listed rates. That calculation does not include every possible operational expense, and the price advantage does not establish that the model will complete a given job successfully or faster.
For teams considering a switch, the practical question is whether 1017 performs reliably in their own tools and tasks. A model tuned for an agent harness may behave differently in another setup. DeepSeek recommends its DeepSeek Harness version 0.2.1-rc.1 or later and says it also tuned the model for Tencent’s CodeBuddy and WorkBuddy.
AI model API cost management tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What DeepSeek’s Chart Shows
DeepSeek’s model card compares 1017 with its September release and four named systems: Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. The new model leads in four of the 11 rows; Claude Opus 5 leads six and GPT-5.6 Sol leads one. Claude Opus 5 has a higher score on the chart’s HLE test, 56.3 compared with 42.6, and on ProgramBench, 37.0 compared with 29.1.
The table’s results are not all directly interchangeable. DeepSeek reports HLE on the full set, while the source notes that its text-only scores are 44.0 for 1017 and 39.1 for the September model; GLM-5.3’s listed 42.0 is a text-only score. Terminal-Bench was run in the Claude Code harness and DeepSWE in Mini-SWE-V2, according to the source report.
The source also describes an early Kingy AI pilot covering seven small tasks, with each task run twice. It says 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs. That limited pilot is not a broad benchmark, and equal scores did not distinguish the models; one reported task took about 27 seconds with 1017 versus about three seconds with Sonnet.
As an affiliate, we earn on qualifying purchases.
Validation and Comparisons Remain Limited
Independent reproduction of DeepSeek’s 11 benchmark results is not provided in the source material. The results should therefore be attributed to DeepSeek, and they do not by themselves establish how the model will perform across different prompts, tool configurations or production workloads.
The comparison chart also excludes newer models named in the source report, including GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 and Gemini 4 Argon. Kingy AI is cited as reporting Terminal-Bench results of 57 to 66 for current frontier models, but those figures come from different evaluations and are not directly comparable to DeepSeek’s chart. The source does not provide enough information to treat that range as a controlled head-to-head result.
DeepSeek has not detailed what changed in training for 1017. The Decoder attributes the improvement to post-training with reinforcement learning using model self-feedback, but that account is not a direct technical explanation from DeepSeek in the supplied material. The model’s reported architecture, parameter counts and cache figures also come from technical materials summarized by third parties, with differing total-size counts noted in the source.
As an affiliate, we earn on qualifying purchases.
Test 1017 Against Your Workload
Developers can now test 1017 through DeepSeek’s website, app or API, where the existing deepseek-flash name routes to the updated version, according to the source. Teams using the older model should verify which version their integration is receiving and compare results on a fixed set of representative tasks before changing production workflows.
A useful evaluation would track successful task completion, retries, latency, token use and the need for human correction—not just benchmark scores. Results should also be checked in the intended agent harness, since DeepSeek says it tuned the model for specific environments. The source material provides no date for an independent benchmark review or further company update, leaving those as open developments.
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 scores on all 11 benchmarks in its model card, with larger gains on Terminal-Bench 4.0 and ProgramBench than on DeepSWE. The company calls it a minor update and has not supplied a detailed account of its training changes in the source material.
Did the API price increase?
No price increase is reported. The source says 1017 replaced the September model at unchanged listed token prices. DeepSeek’s rates vary by input cache status and peak or off-peak hours, so users should check the applicable price schedule for their usage.
Are the benchmark gains independently verified?
Not in the supplied reporting. The 11 benchmark results come from DeepSeek’s own chart. A small Kingy AI pilot is also described, but its seven tasks, each run twice, are too limited to establish broad comparative performance.
Does 1017 outperform every competitor in DeepSeek’s chart?
No. DeepSeek’s chart shows 1017 leading four of 11 benchmark rows. Claude Opus 5 leads six and GPT-5.6 Sol leads one. The chart also does not include several newer models named in the source report.
Should teams switch to 1017 now?
The source does not establish that every team should switch. Developers can compare the updated model with their current system on representative tasks, measuring completion quality, latency, token use and correction needs before deciding.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
