🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: How To Read The Table Before Switching on ThorstenMeyerAI.com
Get the latest gadgets 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 same API name and price. Its published results show gains, especially on some long-running agent tasks, but the benchmarks are DeepSeek-reported, the comparison set omits newer rivals, and independent evidence remains limited.
Users calling the API model name deepseek-flash are now routed to 1017, according to DeepSeek’s release information. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are also routed to the updated model. DeepSeek describes the release as a minor update; it has not published a detailed account of what changed in training.
In the company’s table, Terminal-Bench 4.0 rose from 31.2 to 42.3, a gain of 11.1 points, while ProgramBench increased from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark, moved from 74.2 to 77.0. These are DeepSeek-reported scores; benchmark conditions differ by test, and the table does not establish that every user will see similar improvements on their own workloads.
The model’s weights are available under the MIT license, and DeepSeek lists API input and output rates that are unchanged for this update. The company says its table places 1017 first in four of 11 rows. Claude Opus 5 leads six rows, and GPT-5.6 Sol leads one. The comparison includes selected models, not every current competitor.
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 Tasks
The practical case for switching is not that 1017 tops the table; it does not. It is the combination of reported task improvements and a low API price. DeepSeek’s pricing lists output at $0.60 per million tokens off-peak and $1.20 at peak, with input prices depending on cache use and time. Actual bills depend on token volume, caching and scheduling, so the listed rates alone do not predict the cost of a particular workload.
DeepSeek’s benchmark table suggests the biggest gains are in some longer, multi-step tasks rather than a uniform jump across capabilities. That distinction matters for developers choosing a model for coding agents or automation: a higher score on a particular benchmark may be relevant, but reliability, latency and cost per completed task need testing in the intended setup. The company’s own results are not a substitute for that evaluation.
A limited pilot described by Kingy AI provides some outside evidence, but not a broad validation. It tested seven small tasks twice and reported that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs. Kingy AI said 1017 cost roughly a quarter as much as the other models in that pilot, while also reporting that it was slower on at least one task. The sample is too small to rank the models reliably or establish performance across more demanding work.
As an affiliate, we earn on qualifying purchases.
From September Model to 1017
The release follows DeepSeek’s V4.1 Flash launch on September 10, 2026. The October 17 version takes over the same service names, meaning existing API users may receive it without changing the model identifier. DeepSeek’s published pricing is unchanged, but users should check their own integrations and usage records rather than assume output quality or operating behavior is identical.
DeepSeek describes the system as a sparse Mixture of Experts model with image input and a one-million-token context window. Its technical report, as summarized by Kingy AI, describes a Causal Encoder-Decoder design with 40 layers and 384 routed experts, six active per token, plus a shared expert. Reported model-size figures vary by counting method: the release material gives a 552-billion-parameter backbone, while Hugging Face lists 763 billion. These figures should not be treated as directly comparable without the same accounting basis.
The model card’s rival set includes Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. It does not include newer models named in the source, such as GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon. Kingy AI cites separate Terminal-Bench results of 57 to 66 for current frontier models, but says those evaluations are not directly comparable with DeepSeek’s table. Benchmark version, harness and evaluation setup affect how much weight readers should give any cross-model ranking.
““minor update””
— DeepSeek
As an affiliate, we earn on qualifying purchases.
Independent Results Still Limited
Independent verification remains limited. DeepSeek’s model card supplies the full benchmark table, and the available pilot described by Kingy AI covers only seven small tasks. The source material does not provide independent replications of all 11 benchmarks, or enough detail to establish how the results would transfer to different prompts, tools, codebases or production workloads.
It is also unclear what training or post-training changes produced the gains. The Decoder reported that reinforcement learning with self-feedback may account for them, but DeepSeek has not confirmed that explanation in the supplied material. Details about the impact of the model’s recommended agent harness and compatible tools are also limited. Finally, the company’s comparison does not settle how 1017 performs against newer rivals under matched test conditions.
machine learning benchmark datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Test Before Changing Production
Developers already using deepseek-flash should check whether the automatic update affects their evaluations, latency or output patterns. Before moving production workloads, teams can run a side-by-side test on representative tasks, record pass rates and response times, and compare the full token bill at their expected cache and peak-hour usage. The source material gives no separate migration date because the new version has already replaced the earlier one under the listed names.
Further independent benchmark runs and more detail from DeepSeek on the update’s training and evaluation setup would help clarify the strength of its claims. Until then, the clearest confirmed points are the release date, model-name routing, unchanged listed prices, and the scores DeepSeek published; real-world gains and cost savings remain workload-dependent.
As an affiliate, we earn on qualifying purchases.
Key Questions
What changed with DeepSeek V4.1 Flash 1017?
DeepSeek released 1017 on October 17, 2026, as an update to its September V4.1 Flash model. The company reports higher scores on all 11 benchmarks in its table, with larger gains on Terminal-Bench 4.0 and ProgramBench.
Do users need to change the API model name?
According to DeepSeek’s release information, calls to deepseek-flash now receive 1017. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are also routed to the update.
Is 1017 the top-scoring model in DeepSeek’s table?
No. DeepSeek says 1017 leads four of the 11 listed benchmark rows. Claude Opus 5 leads six, and GPT-5.6 Sol leads one.
Have independent tests confirmed DeepSeek’s benchmark results?
Not across the full benchmark set in the supplied reporting. Kingy AI described a small pilot involving seven tasks run twice, but that sample is not enough to validate all 11 scores or rank the tested models broadly.
Should developers switch immediately?
The update is already served under the existing model names, but whether it suits a particular workload is not established by the company’s table alone. Developers should compare outputs, task success, latency and total token costs on representative work before relying on it in production.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
