DeepSeek V4.1 Flash 1017: Near-Frontier Agent Performance, Lower Cost

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Near-Frontier Agent Performance, Lower Cost on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

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. The company says the update raises scores on all 11 benchmarks in its model card and reduces the cost of some agent tasks; the API price is unchanged, and the reported results have not been independently reproduced across the full benchmark set.

The model is served as deepseek-flash, and DeepSeek has routed the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp to the updated version. The weights are available under the MIT license. The release is described as a minor update to V4.1 Flash, which DeepSeek introduced on September 10.

In DeepSeek’s model-card results, Terminal-Bench 4.0 rose from 31.2 to 42.3, while ProgramBench increased from 20.3 to 29.1. DeepSWE v1.1, another coding benchmark, moved more modestly, from 74.2 to 77.0. These are benchmark score changes reported by DeepSeek, not independent measurements.

DeepSeek’s published comparison lists 1017 ahead on four of 11 benchmarks. Claude Opus 5 leads six rows, including HLE, ProgramBench and NL2Repo-Bench; GPT-5.6 Sol leads one. The benchmark table uses selected competitors, and its scores do not include several newer models named in the source material.

Lower Costs for Long Agent Tasks

The release’s practical case is less about taking the top spot on every benchmark and more about cost per completed task. The Decoder reported DeepSeek’s DeepSWE figures as showing 1017 reaching a 77% score at just under $0.20 per task. A lower-effort setting scored above 72% at about $0.10, while the September model cost about $0.50 per task for a 74.2% score. These figures are from DeepSeek’s benchmark setup, so they should not be treated as a universal price for real-world work.

DeepSeek lists API rates of $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak, rising to $0.30 and $1.20 at peak times. Cached input is priced lower. The company’s stated peak periods are weekdays from 01:00–04:00 and 06:00–10:00 UTC, except Chinese public holidays. That pricing could matter to developers running many long coding or automation jobs, where output volume and repeated context can drive bills.

A small pilot by Kingy AI offers an early, limited check. It reported that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed 14 of 14 runs across seven small tasks tested twice, with 1017 costing roughly a quarter as much. The pilot does not establish a broad ranking: all three models scored perfectly, and at least one task took 1017 about 27 seconds compared with roughly three seconds for Sonnet.

What the Update Changes

DeepSeek describes 1017 as a sparse Mixture of Experts model with image input and a one-million-token context window. Its hosted API supports up to 384,000 output tokens, according to the source material. A technical report summarized by Kingy AI describes a 40-layer Causal Encoder-Decoder design, with 20 encoder and 20 decoder layers, 384 routed experts and six active per token, plus a shared expert.

The model’s size depends on how it is counted: the source gives a 552-billion-parameter backbone, with about eight billion parameters active during prefill and 16 billion during decoding, while Hugging Face lists 763 billion under a different counting method. The report also describes KV-cache compression, with cache use reported at about 890 bytes per token. These technical details describe the system, but DeepSeek has not said what specifically changed in training for the 1017 update.

The Decoder has attributed the gains to post-training with reinforcement learning and self-feedback, in which a model evaluates its outputs using confidence or its own critiques. That account is not a detailed training disclosure from DeepSeek. The company says it tuned the model for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy.

“A “minor update” to V4.1 Flash.”

— DeepSeek, as described in the release material

Benchmark Limits and Open Questions

Independent verification remains limited. The 11 headline benchmark results come from DeepSeek, and the source material says they have not been independently reproduced. The small Kingy AI pilot is separate evidence, but seven tasks tested twice are not enough to settle how the model performs across varied production workloads.

The company’s competitor selection also constrains the comparison. DeepSeek’s chart does not include several newer systems identified in the source, and separately reported Terminal-Bench results for current frontier models—said by Kingy AI to range from 57 to 66—come from different evaluations and are not directly comparable. The size of any performance gap therefore remains uncertain.

It is also unclear which training changes produced the gains. DeepSeek has not provided a specific account of what changed between September and 1017. Cost and speed will vary with task length, cache use, model settings, API demand and whether jobs run during peak pricing hours; benchmark cost-per-task figures do not predict every user’s bill.

Developers Test the New Default

Developers using DeepSeek’s website, app or API will receive 1017 through the updated default model routing, according to the release material. The recommended DeepSeek Harness version is 0.2.1-rc.1 or later; the model is also tuned for Tencent’s CodeBuddy and WorkBuddy. Users who need to assess it for a particular workflow will need to test their own tasks, costs and response times.

The next useful evidence would be independent evaluations using disclosed prompts, settings and comparable model versions, along with more information from DeepSeek about the training changes. The source material also refers to a separate DeepSeek announcement opening search tools to developers, but the available description is incomplete, so its features and timing cannot be reported in detail here.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released V4.1 Flash 1017 on October 17, 2026, replacing the September version on its website, app and API.

Does the update change DeepSeek Flash API prices?

No price change is reported for the update. DeepSeek’s listed token rates vary by input type and time of use, with lower off-peak rates and cheaper cached input.

Are the benchmark gains independently verified?

Not across the full benchmark set. The 11 results are from DeepSeek’s model card; the source material says they have not been independently reproduced. Kingy AI reported a separate, small pilot, but it does not validate all benchmark claims.

Where did 1017 improve most over the September version?

In DeepSeek’s reported scores, Terminal-Bench 4.0 increased by 11.1 points, from 31.2 to 42.3, and ProgramBench rose by 8.8 points, from 20.3 to 29.1.

Does DeepSeek’s chart show 1017 as the top model?

No. DeepSeek’s table has 1017 leading four of 11 rows. Claude Opus 5 leads six and GPT-5.6 Sol leads one, although the chart compares selected models and does not include all newer systems named in the source material.

Source: ThorstenMeyerAI.com