DeepSeek V4.1 Flash 1017: Near-Frontier Agents On A Smaller Budget

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Near-Frontier Agents On A Smaller Budget on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day 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 DeepSeek-V4.1-Flash-1017 on October 17, replacing the September version on its website, app and API while keeping the listed price unchanged. The company reports higher scores on all 11 benchmarks it published, with the largest gains on some multi-step agent tasks; those results are company-reported and have not been independently reproduced.

The update is served through the existing deepseek-flash API name, so users calling that model are routed to 1017 automatically. DeepSeek also routes the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp to the new version. The weights are available under the MIT license, according to the supplied source material.

In DeepSeek’s model-card table, 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 highlighted by DeepSeek, moved from 74.2 to 77.0. These are score changes within DeepSeek’s published evaluation; the source does not establish that the results transfer to every agent system or workload.

DeepSeek says outputs are shorter, which can reduce completion time and token charges. Its published API rates remain $0.15 per million input tokens for uncached off-peak requests and $0.60 per million output tokens, with peak rates twice as high. The price schedule also lists cache-hit rates and specifies peak hours; actual task costs depend on token use and when requests are made.

Lower Costs for Agent Work

The update’s commercial pitch is less about leading every benchmark than about getting useful agent work at a lower cost. DeepSeek’s own figures show notable gains on selected multi-step tasks, and its reported token pricing is low relative to the example prices supplied for some competing models. For developers running many coding or automation jobs, cost per completed task can matter as much as a model’s highest score.

That case remains conditional. A low API bill is useful only if the model completes tasks reliably and the benchmark reflects the work a customer needs done. DeepSeek’s results are not independent validation, and the company’s chart shows 1017 leading four of 11 benchmark rows, not all of them. Claude Opus 5 leads six rows in that chart, including HLE and ProgramBench, while GPT-5.6 Sol leads one.

A small pilot cited by Kingy AI offers an early check, not a broad verdict. Across seven tasks run twice, the pilot reported 14 successful runs out of 14 for 1017, GPT-6.1 Sol and Claude Sonnet 5.5. The models therefore tied in that limited sample, and 1017 was not always fastest. The result supports further testing of cost and reliability, but cannot establish how performance compares across larger or more varied workloads.

The September Model and Its Tests

DeepSeek released the earlier V4.1 Flash version on September 10, 2026; 1017 is described in the source as a minor update. The model is presented as a sparse Mixture of Experts system with image input and a one-million-token context window. The technical details and parameter counts vary by counting method in the cited material, so those figures do not provide a simple measure of user-facing capability.

The benchmark comparisons are DeepSeek’s own and use a selected group of rivals. The source notes that the table compares 1017 with models such as GPT-5.6 Sol and Claude Opus 5, rather than newer versions it names, including GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 and Gemini 4 Argon. Separate Terminal-Bench figures for current frontier models were reported at 57 to 66 by Kingy AI, but the source cautions that those evaluations are not a direct match.

DeepSeek’s table also mixes benchmark conditions: Terminal-Bench was run in the Claude Code harness, while DeepSWE used Mini-SWE-V2. Its HLE result is for the full set; the text-only result is different. Such details affect comparisons, which should be read alongside the scores rather than as a single overall ranking.

“minor update”

— DeepSeek, in its model description as reported by ThorstenMeyerAI.com

Independent Results Still Needed

The main uncertainty is whether DeepSeek’s benchmark improvements will hold up in independent, repeatable testing. The supplied material says no outside reproduction of the published benchmark results was available at release. Kingy AI’s small pilot is a separate test, but its seven-task sample is too limited to settle general performance or model rankings.

DeepSeek has not disclosed in the supplied account what changed in training. The Decoder reports that the gains may come from post-training with reinforcement learning and self-feedback, but that explanation is attributed to the outlet rather than confirmed in the source material as a detailed DeepSeek disclosure. The comparison also leaves open how 1017 performs against newer rival releases under matched test conditions.

The source material introduces a second announcement about DeepSeek search tools called Search and Fetch, but the available text cuts off before describing their functions, availability or terms. Those details cannot be established here.

Testing Across Real Workloads

Developers can now use the updated model through DeepSeek’s website, app and API, or download the MIT-licensed weights, according to the supplied material. The next useful evidence will come from independent evaluations using matched benchmarks, current rival models and clearly reported hardware, harnesses and task conditions.

For teams considering a switch, the practical next step is to test representative jobs and track completion rates, output quality, latency and total token costs—not just benchmark scores. The source also says DeepSeek tuned 1017 for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy, but it does not provide broader deployment results. More information about the announced Search and Fetch tools is also pending in the available account.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released it on October 17, 2026, as an update to the V4.1 Flash model launched on September 10.

Do users need to change the API model name?

No change is indicated for users calling deepseek-flash: the source says that name now routes automatically to 1017. It also says two older model names are routed to the update.

What improved in DeepSeek’s published tests?

DeepSeek reports gains on all 11 benchmarks in its table. Its largest listed increases include Terminal-Bench 4.0, up from 31.2 to 42.3, and ProgramBench, up from 20.3 to 29.1. These results have not been independently reproduced in the supplied material.

Is 1017 proven to outperform frontier models?

No. DeepSeek’s chart shows 1017 leading four of 11 rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. The table uses a selected comparison set and does not include some newer rivals named in the source.

Did DeepSeek change its API prices?

The source says the listed price was unchanged. It gives different rates for cached and uncached input, output and peak versus off-peak periods, so a task’s cost depends on its token use and request timing.

Source: ThorstenMeyerAI.com