🔍 Read the full analysis: The Case For Reviewing DeepSeek V4.1 Flash 1017'S Numbers First on ThorstenMeyerAI.com
Get everyday essentials 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 for its website, app and API users at the same listed price. The company reports gains across 11 benchmarks and lower costs per completed coding task, but the results have not been independently reproduced, and its comparison set does not include newer rivals. An early seven-task pilot supports the cost case but is too small to rank models.
DeepSeek describes 1017 as a minor update to V4.1 Flash, first released on September 10. The default API name deepseek-flash now routes to 1017; the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to it. The model’s weights are available under the MIT license, according to the source material.
In DeepSeek’s model-card table, Terminal-Bench 4.0 rises from 31.2 to 42.3, an 11.1-point gain, while ProgramBench increases from 20.3 to 29.1. Other gains include 7.5 points on Apex Bench and 2.8 points on DeepSWE v1.1. These are score differences between the two versions in the company’s table, not independently verified performance results.
The table also shows that 1017 does not lead every test. It records the top score in four of 11 rows; Claude Opus 5 leads six, and GPT-5.6 Sol leads one. DeepSeek says outputs are shorter, which it presents as a way to finish tasks faster and at lower cost. The source material does not provide an independent measurement of that claim across a broad set of users or workloads.
DeepSeek’s listed API prices are $0.15 per million input tokens for uncached requests off-peak and $0.30 at peak, with output priced at $0.60 and $1.20 per million tokens, respectively. Cached input costs less. The stated peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays; batch jobs outside those windows are priced at half the listed rate, according to the supplied material.
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
What Lower Task Costs Could Change
The release’s practical case is less about taking the top spot on a leaderboard and more about the cost of completing agent work. If DeepSeek’s task-cost figures hold up in independent testing, developers running repeated code and automation jobs could find 1017 economical, particularly when tasks can be scheduled off-peak and use the model’s supported harnesses.
That is a conditional case, not a settled result. The supplied account says DeepSeek’s chart puts 1017 at 77% on DeepSWE for just under $0.20 per task, while a lower-effort setting reached above 72% at about $0.10. It says the September model cost about $0.50 per task for a 74.2% score, and frontier models on that chart cost between $2 and more than $10. Those comparisons are attributed to DeepSeek’s chart as reported by The Decoder; the article material does not establish that all models were tested under identical settings.
For buyers, the distinction matters: a lower token price does not automatically mean a cheaper successful outcome. Performance, retries, latency, task type and integration costs can all affect a real deployment. The numbers are a reason to test the model against a team’s own work, not proof that it will be cheaper or better for every user.
As an affiliate, we earn on qualifying purchases.
How DeepSeek Built Its Comparison
V4.1 Flash is described in the supplied material as a sparse Mixture of Experts model with image input and a one-million-token context window. A technical report, as summarized by Kingy AI, describes a Causal Encoder-Decoder design with 40 layers, 384 routed experts and six active experts per token, plus one shared expert. The material gives different total-size figures depending on how parameters are counted: a 552-billion-parameter backbone and a 763-billion figure listed by Hugging Face. Those figures use different counting conventions.
The report is titled Pushing the Limits of KV Cache Compression; the source says the cache is reported at about 890 bytes per token. The weights are described as an FP8 checkpoint with FP4 experts. DeepSeek has not specified in the supplied account what changed in training. The Decoder reports that reinforcement learning with self-feedback contributed to the gains; that explanation is a reported account, not a training detail confirmed here by DeepSeek.
DeepSeek says 1017 was tuned for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy. Its published comparison names Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. It does not include newer models named in the source material. Kingy AI separately reports Terminal-Bench results of 57 to 66 for current frontier models, but those evaluations are not directly comparable to DeepSeek’s test, so they do not establish a like-for-like ranking.
There is one early outside check in the supplied material. Kingy AI says it ran seven small tasks twice, with 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passing all 14 runs. It found 1017 cost about a quarter as much, but one task took about 27 seconds compared with about three seconds for Sonnet. With all three models scoring perfectly on such a small pilot, the result cannot rank their overall ability.
“Terminal-Bench 4.0: 42.3 for 1017, compared with 31.2 for the September version.”
— DeepSeek, in its model-card benchmark table
As an affiliate, we earn on qualifying purchases.
Which Results Still Need Checking
Independent reproduction is absent for the 11 benchmark results in the supplied material. That leaves open whether other evaluators using different prompts, software versions, harnesses or task sets would see similar gains. DeepSeek’s advantage is also uneven: the strongest improvements in the table are on Terminal-Bench and ProgramBench, while DeepSWE rises by 2.8 points.
The comparison set has another limit: it includes models that the source describes as a generation older than current alternatives. Some separate Terminal-Bench figures for newer models are higher, but they come from different evaluations and cannot be treated as direct comparisons. The exact evaluation settings behind the cost-per-task chart, including how tasks and model configurations were matched, are not fully detailed in the supplied material.
It is also unclear what training changes produced the update. The account attributes self-feedback reinforcement learning to The Decoder, while DeepSeek is said not to specify what changed in training. The Kingy AI pilot is encouraging on its limited tasks, but its seven-task sample is too small to answer questions about reliability across real production workloads, and it found a substantial latency difference on one task.
machine learning performance testing kits
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Deployment Costs
The next useful evidence would be independent, reproducible tests that compare 1017 with current models on the same tasks, harnesses, settings and price assumptions. Such evaluations could show whether the benchmark gains persist outside DeepSeek’s own setup and whether the lower reported cost per task survives when retries and completion quality are counted.
Developers considering the model can test it through the updated deepseek-flash route and compare results with their existing systems, while accounting for token use, peak pricing, latency and any harness requirements. The supplied material recommends DeepSeek Harness version 0.2.1-rc.1 or later. No further release date or independent benchmark schedule is provided, so broader validation remains the next milestone rather than a confirmed forthcoming event.
AI model performance comparison charts
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
When did DeepSeek release V4.1 Flash 1017?
DeepSeek released V4.1 Flash 1017 on October 17, 2026. It replaced the September version on DeepSeek’s website, app and API, according to the supplied material.
Did DeepSeek raise the API price for 1017?
No price increase is reported: the supplied material says the model update came at an unchanged price. It lists different rates for cached and uncached input, output, and peak versus off-peak use.
Does 1017 outperform every model in DeepSeek’s table?
No. DeepSeek’s table gives 1017 the top score in four of 11 benchmarks. Claude Opus 5 leads six rows, and GPT-5.6 Sol leads one.
Have independent evaluators confirmed DeepSeek’s benchmark gains?
The 11 benchmark results in the supplied account are DeepSeek-reported and have not been independently reproduced there. Kingy AI is described as running a small pilot, but that test does not validate the full benchmark table.
What is the strongest case for using 1017?
The case presented is lower cost per completed agent task, alongside gains on some long-running coding and agent benchmarks. The cost figures need broader, matched testing, and actual value will depend on workload, latency and retry rates.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
