DeepSeek V4.1 Flash 1017: The Numbers Behind Its Agent Value

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: The Numbers Behind Its Agent Value on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

DeepSeek released V4.1 Flash 1017 on October 17, replacing the September version across its website, app and API while keeping the listed price unchanged. The company reports higher scores on all 11 benchmarks in its model card, with its largest gains on some long, multi-step agent tasks; the results are DeepSeek’s own figures and have not been independently reproduced across those tests.

The update is served through the deepseek-flash API name, and DeepSeek says older names, including deepseek-v4-flash and deepseek-v4-flash-vision-exp, now route to it. The model’s weights are available under the MIT license on Hugging Face and ModelScope. DeepSeek describes it as a sparse Mixture of Experts model with image input and a one-million-token context window.

In the company’s comparison with its September release, Terminal-Bench 4.0 rose from 31.2 to 42.3, while ProgramBench increased from 20.3 to 29.1. Other changes were smaller: DeepSWE v1.1 rose from 74.2 to 77.0. Across the full table, 1017 has the highest score in four of 11 rows, according to DeepSeek’s model card; Claude Opus 5 leads six, and GPT-5.6 Sol leads one.

DeepSeek’s pricing lists API input at $0.15 per million tokens off-peak when the cache is missed and $0.003 when it is hit; output costs $0.60 per million tokens off-peak. Peak rates are double. The source says output is shorter with 1017, which could reduce both task time and token charges, but provides no independent, broad measurement of those savings.

The Case for Cheaper Agent Tasks

The update’s main commercial proposition is not that 1017 tops every benchmark. It is that developers may be able to run coding and other agent workflows at lower cost per completed task while improving on the previous DeepSeek Flash version. That could matter for teams that run many automated tasks, where cumulative API costs and completion rates can be more relevant than a single leaderboard position.

The Decoder reported DeepSeek’s DeepSWE cost chart as showing 1017 reaching a 77% score at just under $0.20 per task, with a lower-effort setting above 72% at about $0.10. The September model was shown at 74.2% for about $0.50, while frontier models on the chart cost from $2 to more than $10 per task. These are figures from DeepSeek’s evaluation setup as reported by The Decoder, not a general guarantee of cost or performance for other workloads.

Kingy AI described a small pilot involving seven tasks run twice. It reported that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, with 1017 costing roughly a quarter as much as the other systems in that test. The pilot suggests a cost advantage may hold in some tasks, but its small sample cannot rank the three models or establish broad reliability. It also found that 1017 was not always the fastest.

How the October Update Was Measured

DeepSeek calls 1017 a minor update to the V4.1 Flash model released September 10. The company has not detailed what changed in training. The Decoder attributed the gains to post-training with reinforcement learning and self-feedback, but that account is not a detailed public change log from DeepSeek in the supplied material. The company says it tuned the model for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy.

Benchmark comparisons require care. DeepSeek’s table uses selected rivals, including GPT-5.6 Sol and Claude Opus 5, rather than newer models named in the source material. Kingy AI reported Terminal-Bench results of 57 to 66 for current frontier models in separate evaluations; those tests are not directly comparable to DeepSeek’s table. DeepSeek also reports HLE on the full test set for its model, while one listed competitor’s score is text-only, a difference the company’s notes flag.

The model card describes a 552-billion-parameter backbone, with about eight billion active parameters during prefill and 16 billion during decoding. Hugging Face lists 763 billion under a different counting method. Its weights are provided as an FP8 checkpoint with FP4 experts. These technical details and performance figures are vendor-reported; independent replication of the benchmark table is not available in the cited material.

““minor update””

— DeepSeek, describing the release

What the Published Scores Cannot Show

There is no independent reproduction of DeepSeek’s 11 benchmark results in the cited material, and the company has not published details of what changed in training. It is also unclear how reliably the reported task-cost advantage carries over to different agent harnesses, prompts, task types and production workloads. The pilot from Kingy AI is small and does not settle those questions.

The competitor set and evaluation conditions limit direct comparisons. DeepSeek’s table does not include newer rival models cited in the source material, and the separate Terminal-Bench figures for those systems come from different evaluations. Several rows also have missing rival scores, while HLE comparisons mix full-set and text-only results. Readers should treat the table as DeepSeek’s selected comparison, not a complete current ranking.

The source material also refers to a separate announcement opening DeepSeek search tools to developers, but its description ends before explaining how the tools work, their availability or any pricing. Those details cannot be confirmed from the supplied information.

Questions for Wider Testing

The next useful evidence would be independent benchmark runs that disclose the model version, harness, prompts, scoring method and cost accounting. Repeated tests across longer and more varied agent tasks could show whether the gains on Terminal-Bench and ProgramBench translate into more dependable task completion, and whether shorter outputs consistently lower total spending.

Developers can access 1017 through DeepSeek’s hosted services and download the open-licensed weights, while the company recommends its DeepSeek Harness version 0.2.1-rc.1 or later. The source provides no schedule for a further model update or for additional independent evaluations. Any conclusions about its standing against newer competitors remain provisional until comparable tests are published.

Key Questions

What changed in DeepSeek V4.1 Flash 1017?

DeepSeek says the October 17 update improves results on all 11 benchmarks in its model card compared with the September release. Its largest listed gains include Terminal-Bench 4.0, up from 31.2 to 42.3, and ProgramBench, up from 20.3 to 29.1.

Did DeepSeek raise its API prices?

No price change was reported with the update. DeepSeek’s listed off-peak rates are $0.15 per million input tokens when the cache is missed, $0.003 per million cached input tokens and $0.60 per million output tokens; peak rates are twice those amounts.

Has an independent evaluation confirmed DeepSeek’s benchmark gains?

Not in the cited material. The 11 benchmark figures come from DeepSeek’s model card. Kingy AI reported a separate, small pilot across seven tasks run twice, but that does not independently reproduce the full benchmark table.

Is 1017 the top model in DeepSeek’s comparison?

No. DeepSeek’s table gives 1017 the highest score in four of 11 rows. Claude Opus 5 leads six rows and GPT-5.6 Sol leads one, according to the company’s figures.

What remains unknown about the update?

DeepSeek has not detailed the training changes behind 1017, and the reported benchmark and cost results have not been broadly reproduced independently. The source also does not provide the full details of the separate search-tools announcement.

Source: ThorstenMeyerAI.com