DeepSeek V4.1 Flash 1017: Near-Frontier Agents For A Tenth The Price

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Near-Frontier Agents For A Tenth The Price on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get movie nights 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 token prices unchanged. The company reports improvements on all 11 benchmarks it published, with its largest gains on agent tasks that involve extended, multi-step work; those results remain company-reported and not independently reproduced.

The update is served as deepseek-flash; DeepSeek also routes the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp to the new version. The model’s 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 the company’s model card, Terminal-Bench 4.0 rose from 31.2 to 42.3, and ProgramBench from 20.3 to 29.1. The DeepSWE v1.1 score moved from 74.2 to 77.0, a smaller gain. DeepSeek says outputs are shorter, which can help tasks finish faster and lower their cost. It has not disclosed what changed in training. The Decoder reported that the improvement may involve post-training with reinforcement learning and self-feedback, but that account is not a detailed training disclosure from DeepSeek.

DeepSeek’s cost case rests on the price per completed task, not a lower token price. The company lists peak-hour output at $1.20 per million tokens and off-peak output at $0.60; peak input without a cache hit is $0.30 per million tokens, compared with $0.15 off-peak. In a DeepSWE comparison reported by The Decoder, 1017 reached 77% at just under $0.20 per task. The September version reached 74.2% at about $0.50 per task. These figures come from DeepSeek’s evaluation and do not establish performance across every workload.

Lower Costs for Long Agent Tasks

The update could make automated coding and other agent workloads cheaper for developers who already use DeepSeek’s API or run its open weights. At DeepSeek’s listed prices, one illustrative request with one million uncached input tokens and 200,000 output tokens costs $0.27 off-peak. The same token volumes at $2 per million input tokens and $10 per million output tokens would cost $4. Actual bills depend on cache use, time of day, and token volumes.

An early Kingy AI pilot offers limited outside evidence. Across seven small tasks run twice, 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs; Kingy AI estimated that 1017 cost roughly one quarter as much as the other models. The pilot is too small to rank the systems, and one task took about 27 seconds on 1017 versus about three seconds on Sonnet. It suggests a possible cost advantage for those tasks, while showing that lower cost does not mean faster completion in every case.

The release also gives developers a new reason to weigh task success, latency and total cost together. A model that uses fewer tokens or costs less per run can still be a poor fit if it needs more retries, takes longer, or fails on a team’s particular work. DeepSeek’s published comparison does not settle those questions across real production workloads.

How the Benchmark Compares

DeepSeek compares 1017 with its September Flash version and four other models in its own chart. It reports 1017 leading four of 11 rows; Claude Opus 5 leads six, and GPT-5.6 Sol leads one. Opus 5 scores 56.3 on the chart’s full-set HLE reasoning benchmark, against 42.6 for 1017, and leads on ProgramBench and NL2Repo-Bench. The results do not show 1017 as the top scorer overall.

The chart also uses models that do not represent every current frontier release. It includes GPT-5.6 Sol and Claude Opus 5, rather than newer systems named in the source material. Kingy AI cited Terminal-Bench figures of 57 to 66 for current frontier models from separate evaluations. Because those runs are not a direct, controlled comparison with DeepSeek’s chart, they provide context but cannot be treated as equivalent scores.

Some benchmark results depend on evaluation setup. DeepSeek’s HLE figures use the full set; its text-only score for 1017 is 44.0, while the listed GLM-5.3 score of 42.0 is text-only. Terminal-Bench used the Claude Code harness, and DeepSWE used Mini-SWE-V2. DeepSeek says the model is tuned for its own harness and named Tencent tools, so performance may vary with the software environment.

Independent Results Remain Limited

DeepSeek’s benchmark results have not been independently reproduced at scale. The Kingy AI pilot covered only seven small tasks, and all three tested models passed every run, leaving little basis to compare their capabilities. It also does not establish whether the cost difference holds for larger or more varied workloads.

DeepSeek has not explained the specific training changes behind the update. The report linking the gains to reinforcement learning with self-feedback is attributed to The Decoder, and the method has not been confirmed in a detailed disclosure by DeepSeek in the material available here. The relationship between the model’s stated size and other published counts is also unclear: the source describes a 552-billion-parameter backbone, while Hugging Face lists 763 billion using a different counting method.

The benchmark chart’s chosen rivals and evaluation setups limit direct comparisons with newer models. The source also ends partway through its description of DeepSeek’s newly opened search tools, so the tools’ full capabilities, access terms and availability are not established here.

Search Tools and Wider Testing

Developers can use the updated model through DeepSeek’s website, app and API, with calls to deepseek-flash routed to 1017. Its open MIT-licensed weights are available on Hugging Face and ModelScope. Independent testing across larger task sets, different agent harnesses and real production workloads would help establish whether the reported gains and lower task costs hold beyond the company’s benchmarks.

DeepSeek has also opened its search engine to developers as two agent tools, according to the source report. The available material begins to describe a real-time search tool but cuts off before giving the tools’ complete behavior or terms. Further details about their operation and access remain the next concrete development to watch.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released it on October 17, 2026, replacing the September Flash model on its website, app and API.

Did DeepSeek lower the model’s token prices?

No. DeepSeek kept its listed token prices unchanged. The company says shorter outputs can reduce the cost per completed task, depending on the workload.

Which tasks showed the largest reported gains?

DeepSeek reported its largest listed increases on Terminal-Bench 4.0, up from 31.2 to 42.3, and ProgramBench, up from 20.3 to 29.1. These are company-reported benchmark results.

Have independent tests confirmed DeepSeek’s benchmark claims?

Not at scale. Kingy AI reported a pilot covering seven small tasks, with 14 successful runs each for 1017, GPT-6.1 Sol and Claude Sonnet 5.5. That small test does not independently reproduce DeepSeek’s full benchmark results.

What else did DeepSeek announce?

DeepSeek opened its search engine to developers through two agent tools. The source material does not provide enough detail to describe their full features or access terms.

Source: ThorstenMeyerAI.com