DeepSeek V4.1 Flash 1017: Weighing Cost Against Agent Results

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Weighing Cost Against Agent Results on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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 on its website, app and API while keeping listed token prices unchanged. The company says the update improves results on all 11 benchmarks in its model card and reduces the cost of some agent tasks; those benchmark results have not been independently reproduced across the full set.

The updated model is served as deepseek-flash, and DeepSeek says requests using that name now reach version 1017 automatically. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to the update. DeepSeek describes the release as a minor update to V4.1 Flash, first released on September 10, 2026. Its weights are available under the MIT license, according to the supplied source material.

In DeepSeek’s model card, 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 score changes reported by DeepSeek; the source does not establish that an independent evaluator has replicated the full benchmark table. DeepSeek says outputs are shorter, which can make tasks faster and cheaper, but the source does not provide a general measured time saving across workloads.

The company lists API prices of $0.15 per million input tokens for uncached off-peak use and $0.30 at peak, with output priced at $0.60 and $1.20 per million tokens, respectively. Cache-hit input costs $0.003 off-peak and $0.006 at peak. DeepSeek defines weekday peak hours as 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays; batch jobs outside those hours are listed at half price. Prices and task costs depend on token use and the work being attempted.

Lower Costs for Agent Work

The release puts the focus on cost per completed task, a practical measure for developers running coding agents repeatedly. DeepSeek’s chart, as reported by The Decoder, puts 1017 at 77.0% on DeepSWE for just under $0.20 per task. A low-effort setting reportedly scores above 72% at about $0.10, while the September model cost about $0.50 per task for a 74.2% score. These figures come from the reported chart and may depend on its evaluation setup.

DeepSeek’s benchmark table does not show 1017 leading every category. It leads four of 11 rows; Claude Opus 5 leads six, including HLE and ProgramBench, while GPT-5.6 Sol leads one. That distinction matters for teams choosing a model: a lower reported task cost may be useful even when another model scores higher, but performance can vary by task, tool setup and evaluation method.

The available comparison also has a time limitation. DeepSeek’s chosen rivals include GPT-5.6 Sol and Claude Opus 5, while the source says the table omits newer models. Kingy AI separately reported Terminal-Bench results between 57 and 66 for current frontier models, but those evaluations differ and are not a direct comparison with DeepSeek’s table. The figures suggest a gap may remain, though they do not settle how the models compare under one shared test.

From September Model to 1017

DeepSeek introduced the previous V4.1 Flash model on September 10, 2026. The October 17 release replaces it under the same API model name, so users calling deepseek-flash receive 1017 without changing that name. The listed token prices also remain unchanged, according to the source material.

The model is described as a sparse Mixture of Experts system with image input and a one-million-token context window. A technical report summarized by Kingy AI describes a Causal Encoder-Decoder design with 40 layers and 384 routed experts, six active per token, plus one shared expert. The source reports about 890 bytes of key-value cache per token. Model size is described differently across sources: the report gives a 552-billion-parameter backbone, while Hugging Face lists 763 billion using a different count. These figures are not directly interchangeable.

DeepSeek recommends its DeepSeek Harness at version 0.2.1-rc.1 or later and says it tuned 1017 for that environment as well as Tencent’s CodeBuddy and WorkBuddy. The Decoder attributes the reported gains to post-training with reinforcement learning and self-feedback. DeepSeek has not specified what changed in training, according to the supplied material, so the cause of individual benchmark gains remains unconfirmed.

“minor update”

— DeepSeek, describing the October release

Benchmark Evidence Still Limited

Most benchmark figures come from DeepSeek, and the supplied source says they have not been independently reproduced as a full set. The small Kingy AI pilot offers a limited outside check: across seven tasks run twice, DeepSeek 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs. With all three scoring perfectly, the pilot cannot establish which is more reliable overall.

That pilot also found that 1017 was not always fastest: one task took about 27 seconds, compared with about three seconds for Sonnet. The source says 1017 marked unsupported facts as unknown and ignored instructions inserted into input, but those observations cover only the tested tasks. It remains unclear how results would change across broader workloads, different agent harnesses, longer runs or a common evaluation using current rival models.

The source material ends while introducing a separate DeepSeek search announcement, so it does not provide enough detail to describe the search tools or assess their availability. It also does not document DeepSeek’s training changes directly. Those points should not be inferred from the benchmark gains.

Broader Testing and Tool Details

The next useful evidence would be independent, reproducible evaluations that publish task definitions, harnesses, token use, completion rates and costs for 1017 and current alternatives under the same conditions. Such comparisons could clarify whether the reported savings hold across workloads, and whether shorter outputs consistently reduce end-to-end cost without lowering task quality.

Developers can access the update through DeepSeek’s website, app and API, according to the source. Those evaluating it should account for peak and off-peak pricing, cached versus uncached input, and the token volume their tasks generate. The source also says DeepSeek recommends its own harness and lists support for Tencent’s CodeBuddy and WorkBuddy, but gives no dates or further rollout details for those integrations.

More information is needed about the separately announced search tools and the training changes behind 1017. Until those details and broader outside results are available, the clearest confirmed development is the model replacement and the unchanged listed prices; performance and cost advantages remain claims that need wider testing.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released it on October 17, 2026, replacing the September V4.1 Flash version on its website, app and API, according to the supplied source.

Did DeepSeek change its listed API prices?

No price change is reported. The source says the model update went live at unchanged prices. Input and output rates vary by cache status and by whether usage falls in DeepSeek’s stated peak hours.

Which benchmarks improved most over the September model?

In DeepSeek’s table, Terminal-Bench 4.0 rose by 11.1 points, from 31.2 to 42.3, and ProgramBench rose by 8.8 points, from 20.3 to 29.1. These are company-reported scores, not independently replicated results in the supplied material.

Has an independent test confirmed DeepSeek’s results?

The supplied source says the full benchmark set has not been independently reproduced. Kingy AI reported a small pilot in which 1017 and two other models each passed 14 of 14 runs, but that test was too small to rank the models.

What remains unknown about the update?

DeepSeek has not specified what changed in training, according to the source. Wider performance across tasks and agent setups, comparable results against newer models, and details about the separately mentioned search tools remain unclear.

Source: ThorstenMeyerAI.com