DeepSeek V4.1 Flash 1017 Puts Near-Frontier Agents Within Reach

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017 Puts Near-Frontier Agents Within Reach 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 DeepSeek-V4.1-Flash-1017 on October 17, replacing its September Flash model on the company’s website, app and API. DeepSeek says the update improves performance on its published benchmarks and produces shorter outputs, which can make agent tasks faster and cheaper; the hosted model’s listed token prices are unchanged.

The release is served as deepseek-flash; DeepSeek also routes the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp to the updated model. The weights are available under the MIT license. DeepSeek describes the release as a minor update and recommends its DeepSeek Harness version 0.2.1-rc.1 or later, as well as Tencent’s CodeBuddy and WorkBuddy.

In DeepSeek’s model-card table, Terminal-Bench 4.0 rises from 31.2 to 42.3, while ProgramBench moves from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark, increases by 2.8 points, from 74.2 to 77.0. These are scores reported by DeepSeek, not results independently reproduced across the full benchmark set.

The model card lists an input context of one million tokens and up to 384,000 output tokens on the hosted API. The technical report describes a sparse Mixture of Experts design. DeepSeek’s price list sets output at $0.60 per million tokens off-peak and $1.20 at peak, with lower rates for cached input. The company says it has made outputs shorter; it does not publish a change to the per-token price.

Lower Costs for Long Agent Runs

The release’s clearest potential appeal is cost per completed task, particularly for developers running coding or other multi-step agents repeatedly. Token prices alone do not determine total cost: the number of tokens used, retries, task success and any additional tools all matter. DeepSeek says shorter outputs reduce usage, but customers will need to measure their own workloads to see whether that translates into savings.

On DeepSeek’s reported DeepSWE cost chart, 1017 reaches a 77.0% score at just under $0.20 per task; a low-effort setting is reported above 72% at about $0.10. The September model is shown at 74.2% for about $0.50 per task. These figures come from the company’s evaluation and do not establish the cost or success rate for every real-world agent workflow.

A small pilot by Kingy AI offers an initial outside data point. Across seven tasks run twice, it reports that 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, with 1017 costing roughly a quarter as much. The sample is too small to rank the systems, and the report says 1017 was not always fastest. For teams choosing a model, the results make a case for testing cost and latency alongside benchmark scores.

A Narrow Benchmark Comparison

DeepSeek compares the October model with its own September release and four selected competitors: Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. The company’s table shows 1017 leading four of 11 rows. Claude Opus 5 leads six, including HLE, ProgramBench, NL2Repo-Bench and two of the three vision tests; GPT-5.6 Sol leads the remaining row.

The benchmark results have important qualifications. DeepSeek says its HLE figures use the full set: 42.6 for 1017 and 36.8 for the September model. Its text-only figures are 44.0 and 39.1 respectively, while the table’s GLM-5.3 score of 42.0 is text-only. Terminal-Bench was run in the Claude Code harness and DeepSWE in Mini-SWE-V2, according to the source material.

The comparison does not include several newer models named in the source material. Kingy AI has reported Terminal-Bench results between 57 and 66 for current frontier models, but those figures came from separate evaluations and are not a direct comparison with DeepSeek’s table. DeepSeek’s benchmarks therefore show progress against its chosen set, not a settled ranking against every current model.

““minor update””

— DeepSeek

Independent Results Remain Limited

Most of the benchmark evidence comes from DeepSeek, and the source material does not report independent reproduction of the full table. The Kingy AI pilot is small: seven tasks, each run twice. Its equal pass counts cannot show which model is more reliable across a wider range of work.

It is also unclear what changed in training. DeepSeek has not specified the changes, while The Decoder reports that the gains came from post-training with reinforcement learning and self-feedback. That account should be treated as a reported explanation, not as a detailed method confirmed by DeepSeek in the supplied material.

The source material begins describing a separate announcement: DeepSeek opened search tools for developers, with a Search tool said to return real-time results. The account cuts off before giving the tools’ full details, so their features, access terms and relationship to 1017 cannot be established here.

Testing in Developer Workflows

Developers can try 1017 through DeepSeek’s hosted services or use the MIT-licensed weights. A useful next check is whether the reported gains hold on varied, independently run tasks and whether shorter outputs reduce end-to-end cost without lowering success rates. Latency, retries and tool use should be measured alongside token spend.

Further details on DeepSeek’s training changes and its announced developer search tools would help clarify the scope of the update. Until broader evaluations appear, the October release is best understood as a company-reported improvement with an early, limited outside pilot—not a definitive comparison across the model market.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek shipped it on October 17, 2026. It replaced the September version on the company’s website, app and API.

Did DeepSeek raise the model’s API prices?

No price increase is reported. The supplied price list gives output rates of $0.60 per million tokens off-peak and $1.20 at peak, with separate lower rates for cached input.

Which benchmark results improved most?

In DeepSeek’s table, Terminal-Bench 4.0 rose from 31.2 to 42.3, and ProgramBench rose from 20.3 to 29.1. The figures are reported by DeepSeek and have not been independently reproduced across the full set.

Is 1017 the top model in DeepSeek’s comparison?

No. DeepSeek’s table has 1017 leading four of 11 rows. Claude Opus 5 leads six, and GPT-5.6 Sol leads one.

How much independent testing is available?

The supplied material describes a Kingy AI pilot covering seven tasks run twice. Three models passed all 14 runs, so the pilot offers an early cost and performance signal but cannot establish a broad ranking.

Source: ThorstenMeyerAI.com