How DeepSeek V4.1 Flash 1017 Brings Near-Frontier Agents To A Lower Cost

🔍 Read the full analysis: How DeepSeek V4.1 Flash 1017 Brings Near-Frontier Agents To A Lower Cost on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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 token prices unchanged. The company says the update improves performance across all 11 benchmarks it published, with its largest gains on some multi-step agent tasks; the results have not been independently reproduced across the full benchmark set.

Calls to deepseek-flash now use version 1017 automatically, and the older aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to the new release. DeepSeek describes the update as minor. Its model card reports that 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, rose more modestly from 74.2 to 77.0.

The company says version 1017 produces shorter outputs, which can reduce completion time and token charges. DeepSeek’s published token rates are $0.15 per million uncached input tokens off-peak and $0.30 at peak, with output priced at $0.60 and $1.20, respectively. Cached input costs less. At the off-peak rates, a workload using one million uncached input tokens and 200,000 output tokens would cost $0.27, according to the stated price schedule.

The model’s weights are available under the MIT license on Hugging Face and ModelScope. DeepSeek lists a one-million-token context window and up to 384,000 output tokens on its hosted API. The technical report describes a sparse mixture-of-experts model; it does not explain what changed in training for this update. The Decoder attributes the reported gains to post-training with reinforcement learning and self-feedback, but that description is not confirmed in the source material by DeepSeek’s model card.

Lower Prices Could Expand Agent Use

The release makes the case for DeepSeek Flash on cost per completed task, rather than a claim that it leads overall model rankings. The Decoder reported DeepSeek’s chart showing version 1017 reaching 77% on DeepSWE at just under $0.20 per task; a low-effort setting exceeded 72% at about $0.10. On that same chart, the September model cost about $0.50 per task for a 74.2% score, while the compared frontier models cost between $2 and more than $10. These are figures reported from DeepSeek’s evaluation, not a broad independent cost study.

Lower costs could make it practical for developers to run more coding or workflow tasks through an agent, especially when those jobs consume many tokens. The savings depend on the task, output length, cache use and whether requests fall in peak pricing hours. Peak periods are listed as 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays; batch jobs outside those hours are charged at half the peak rate.

Kingy AI reported an early pilot of seven small tasks run twice: version 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, at roughly a quarter of the other models’ cost. That pilot offers limited support for the cost argument, but its small task set cannot establish broad performance or rank models that all scored perfectly. Kingy AI also reported that DeepSeek was not always fastest: one task took about 27 seconds, compared with about three seconds for Sonnet.

A Small Update With Broad Routing

DeepSeek released the prior V4.1 Flash version on September 10, 2026. The October update replaces that version for users of the company’s site, app and API, so users of the default alias receive 1017 without changing the model name in their requests. DeepSeek recommends its own DeepSeek Harness version 0.2.1-rc.1 or later and says the model is tuned for that harness and Tencent’s CodeBuddy and WorkBuddy.

DeepSeek’s benchmark table compares version 1017 with its September predecessor and four selected rivals. It leads in four of the 11 rows; Claude Opus 5 leads in six, and GPT-5.6 Sol leads in one. Claude Opus 5 scores 56.3 on the table’s full-set HLE measure, compared with 42.6 for 1017, and 37.0 on ProgramBench, compared with 29.1. DeepSeek notes that its HLE figures use the full set, while one rival’s listed figure uses a text-only subset, so those particular numbers do not share the same basis.

The comparison also does not include newer models named in the source material, including GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 and Gemini 4 Argon. Kingy AI cites separate Terminal-Bench 4.0 figures of 57 to 66 for current frontier models, but those came from different evaluations and are not directly comparable with DeepSeek’s table.

““minor update””

— DeepSeek

Independent Results Remain Limited

Most benchmark figures come from DeepSeek, and the source material says none has been independently reproduced across the full suite. The company has not detailed what changed in training. The reported cost-per-task results depend on DeepSeek’s evaluation setup and may not predict costs or success rates for other agent harnesses and real-world tasks.

The selected comparison set leaves out newer competing models, and some listed benchmark scores use different subsets or evaluation setups. DeepSeek’s table says Terminal-Bench ran in the Claude Code harness and DeepSWE in Mini-SWE-V2. The supplied source also begins describing a separate announcement of search tools but cuts off before details are provided, so their features and availability cannot be established here.

API Use Will Test the Claims

Developers can access version 1017 through DeepSeek’s existing site, app and API, where the default deepseek-flash alias now routes to the update. Its practical performance will become clearer as users test it on their own tasks and independent evaluators reproduce the benchmark and cost figures using comparable methods.

For now, the next evidence to watch is whether independent tests confirm the reported gains on long agent runs, and whether the lower cost per task holds across different workloads. DeepSeek has not provided a training-change breakdown in the supplied material, and no further release date or evaluation milestone is specified.

Key Questions

What changed in DeepSeek V4.1 Flash 1017?

DeepSeek says version 1017 improves scores on all 11 benchmarks in its model card, with larger reported gains on Terminal-Bench 4.0 and ProgramBench. The update replaced the September version on DeepSeek’s website, app and API on October 17, 2026.

Did DeepSeek raise its API prices?

No price change is reported. DeepSeek’s listed rates vary by peak hours and whether input tokens are cached; the off-peak price for uncached input is $0.15 per million tokens, and output is $0.60 per million.

Are the benchmark gains independently verified?

The supplied source says DeepSeek’s full benchmark results have not been independently reproduced. Kingy AI reported a small pilot covering seven tasks run twice, but that sample cannot establish broad performance.

Does version 1017 outperform the rival models in DeepSeek’s chart?

It leads four of the 11 benchmark rows in DeepSeek’s table. Claude Opus 5 leads six and GPT-5.6 Sol leads one. The table does not include several newer rivals, and some comparisons use different evaluation subsets or setups.

How can developers access the update?

DeepSeek routes the deepseek-flash API alias to version 1017. The previous aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp also route to it, according to the supplied source.

Source: ThorstenMeyerAI.com