DeepSeek V4.1 Flash 1017 Explained: Agents, Pricing, And Tradeoffs

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017 Explained: Agents, Pricing, And Tradeoffs 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 on its website, app and API. The company says the update improves performance on all 11 benchmarks it published, with its largest reported gains on certain multi-step agent tasks; the listed token prices are unchanged.

DeepSeek says Terminal-Bench 4.0 rose from 31.2 to 42.3, while ProgramBench increased from 20.3 to 29.1. Its DeepSWE score rose more modestly, from 74.2 to 77.0. These are scores reported in DeepSeek’s model card, not independently reproduced results. The company also says outputs are shorter, which can reduce completion time and expense, but the supplied material does not give a broad independent measurement of that claim.

The update is served under the API name deepseek-flash; older names, including deepseek-v4-flash, are also routed to the new version. DeepSeek says requests using those names receive 1017 automatically. Its weights are available under the MIT license, and the report describes a sparse mixture-of-experts model with image input and a one-million-token context window. The hosted API is listed as supporting up to 384,000 output tokens.

DeepSeek’s published pricing is $0.15 per million uncached input tokens off-peak and $0.30 at peak, with cached input at $0.003 and $0.006, respectively. Output costs $0.60 off-peak and $1.20 at peak per million tokens. The company defines weekday peak periods as 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. Batch requests scheduled outside peak periods are listed at half price.

Lower Costs for Agent Work

The update’s main practical case is not that DeepSeek tops every benchmark; it is that some coding and agent tasks may be completed at lower cost. A cost comparison cited by The Decoder says 1017 reached 77% on DeepSWE at just under $0.20 per task, while a lower-effort setting scored above 72% at about $0.10. Those figures come from a benchmark chart and should not be treated as a guaranteed cost for other workloads.

At the stated off-peak prices, a job using one million uncached input tokens and 200,000 output tokens would cost about $0.27, according to the source material. That makes the model relevant to developers running repeated or long agent jobs, where usage charges can accumulate. Actual bills will depend on cache use, timing, token volume and the task itself.

There is also a reason to read the performance claims cautiously. DeepSeek’s chart shows 1017 leading four of 11 rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. The results indicate progress over DeepSeek’s previous version, but do not establish that it is the strongest available model or that every agent workflow will be cheaper or more reliable.

A Replacement for September Flash

DeepSeek calls 1017 a minor update to V4.1 Flash, which it released on September 10. The new model replaces that release on the company’s consumer products and API, rather than appearing as a separate default offering. The source says users calling the generic deepseek-flash name are switched to 1017 automatically, with no change to listed prices.

The technical report, as summarized by Kingy AI, describes a Causal Encoder-Decoder architecture with 40 layers and 384 routed experts, six active per token, alongside a shared expert. The report gives a 552-billion-parameter backbone, with about eight billion active parameters during prefill and 16 billion during decoding; Hugging Face lists 763 billion using a different counting method. Those counts are not directly interchangeable. The report also focuses on key-value cache compression, with a reported cache size of about 890 bytes per token.

DeepSeek has not detailed what changed in training for this release. The Decoder reports that reinforcement learning with self-feedback contributed to the gains, but that explanation is attributed to the outlet’s account, not a full training-change disclosure from DeepSeek. The company says 1017 was tuned for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy.

“minor update”

— DeepSeek, in its description of the release

Independent Tests Still Needed

Independent reproduction of DeepSeek’s benchmark table is not provided, and the source notes that its chosen rivals are older than models now available. Separate Terminal-Bench results for newer frontier systems are reported in the range of 57 to 66 by Kingy AI, but those evaluations differ and are not a direct comparison with DeepSeek’s chart.

Kingy AI’s pilot offers limited outside evidence, not a broad validation: it covered seven small tasks, run twice each, and three models achieved perfect results. The pilot also found that 1017 was not always the fastest; one task reportedly took about 27 seconds, compared with roughly three seconds for Sonnet. The material does not establish how representative those tasks are of production workloads.

It also remains unclear exactly which training changes produced the gains, how performance varies across longer real-world jobs, and how costs compare across different agent setups. DeepSeek’s report does not provide enough information here to settle those questions.

Testing in Real Agent Workflows

Developers can access 1017 through DeepSeek’s site, app and API, or download the MIT-licensed weights from Hugging Face and ModelScope. DeepSeek recommends DeepSeek Harness version 0.2.1-rc.1 or later for its own agent setup. Broader use should show whether the benchmark improvements translate into dependable results across different tools and tasks.

The next useful evidence will be independently run evaluations using current rival models, disclosed task settings and comparable pricing assumptions. Until those results are available, the clearest confirmed development is the replacement of the September model and DeepSeek’s unchanged price schedule; the performance and cost advantages remain claims to test across real workloads.

Key Questions

What changed with DeepSeek V4.1 Flash 1017?

DeepSeek says the October 17 update improves its scores across 11 published benchmarks, including Terminal-Bench 4.0 and ProgramBench. It replaces the September version on DeepSeek’s website, app and API.

Did DeepSeek raise its API prices?

No price increase is reported. DeepSeek’s listed rates remain $0.15 per million uncached input tokens off-peak and $0.60 per million output tokens off-peak, with higher peak rates and lower cached-input rates.

Are the benchmark gains independently verified?

The supplied source material does not include independent reproduction of DeepSeek’s 11-benchmark table. Kingy AI reported a small pilot, but it covered only seven tasks run twice each and cannot establish broad performance.

Is 1017 the top-performing model in DeepSeek’s comparison?

No. DeepSeek’s chart shows 1017 leading four of 11 benchmark rows. Claude Opus 5 leads six, and GPT-5.6 Sol leads one; the chart also does not include newer rivals listed in the source material.

Can developers use the model’s weights?

Yes. The source says the weights are available under the MIT license on Hugging Face and ModelScope. DeepSeek also serves the model through its website, app and API.

Source: ThorstenMeyerAI.com