DeepSeek V4.1 Flash 1017: Near-Frontier Agent Work At Lower Prices

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: Near-Frontier Agent Work At Lower Prices 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 Oct. 17, replacing the September version across its website, app and API while keeping the listed token prices unchanged. The company reports gains on all 11 benchmarks in its model card, with larger improvements on some multi-step agent tasks, but those results have not been independently reproduced across the full benchmark suite.

DeepSeek’s model card reports Terminal-Bench 4.0 rising from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1 compared with the September V4.1 Flash. DeepSWE v1.1, a coding benchmark, increased more modestly, from 74.2 to 77.0. These are scores reported by DeepSeek, not independent evaluations.

The company says shorter outputs make tasks faster and cheaper to complete. The Decoder reported that, on DeepSeek’s DeepSWE cost chart, 1017 reached a 77% score at just under $0.20 per task; a lower-effort setting scored above 72% at about $0.10. The September model was shown at about $0.50 per task for a 74.2% score. These task-cost figures depend on the evaluation setup and should not be treated as universal costs.

The API price list gives separate rates for peak and off-peak periods. Per million tokens, input costs $0.15 off-peak when cache misses and $0.003 for cache hits; output costs $0.60 off-peak. Peak rates are twice those amounts. DeepSeek routes the `deepseek-flash` name to 1017, and the older names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` are also routed to it. The weights are available under the MIT license, according to the source material.

Lower Costs for Agent Tasks

The release’s main pitch is cost per completed task, not a claim that 1017 leads every performance table. For developers running agents repeatedly—such as coding tools that inspect a repository, make edits and test the result—lower token prices and shorter outputs could reduce operating expenses. That matters even if a model’s raw benchmark score trails more expensive competitors.

DeepSeek’s own chart shows a mixed competitive picture. Version 1017 leads in four of the 11 listed rows, while Claude Opus 5 leads six and GPT-5.6 Sol leads one. Opus 5 scores higher on several reasoning, coding and vision measures. This suggests the update may be useful for particular agent workloads without establishing that it is a general replacement for higher-ranked models.

Kingy AI reported an early pilot covering seven small tasks, each run twice. It said 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs, with 1017 costing roughly a quarter as much. The pilot also noted that 1017 was not always fastest: one task took about 27 seconds, compared with about three seconds for Sonnet. The sample is too small to determine broad reliability, speed or relative quality.

What DeepSeek Changed in 1017

DeepSeek described the release as a minor update to V4.1 Flash, which it released on Sept. 10. The company has not specified what changed in training. The Decoder reported that the improvement may involve post-training with reinforcement learning and self-feedback, in which a model evaluates its own outputs. That explanation is a report about the approach, not a training account confirmed in detail by DeepSeek.

According to the technical report as summarized by Kingy AI, the model uses a sparse mixture-of-experts design and accepts images. The report describes a one-million-token context window and a Causal Encoder-Decoder architecture. The source material gives differing parameter totals: a 552-billion-parameter backbone with about 8 billion active during prefill and 16 billion during decoding, while Hugging Face lists 763 billion using a different count. Those figures are not directly interchangeable.

DeepSeek recommends its own DeepSeek Harness, version 0.2.1-rc.1 or later, and says the model is tuned for Tencent’s CodeBuddy and WorkBuddy. In its benchmark table, Terminal-Bench was run in the Claude Code harness and DeepSWE in Mini-SWE-V2. Harness and evaluation choices can affect results, so scores across different setups need care.

““minor update””

— DeepSeek, describing the release

Benchmark Claims Still Need Testing

Independent validation remains limited. The 11 benchmark results were published by DeepSeek, and the supplied material does not show an independent reproduction of that full table. Kingy AI’s pilot is small, covering only seven tasks run twice, so it cannot establish performance across varied production workloads.

The comparison set also leaves out newer competitors. DeepSeek’s table includes GPT-5.6 Sol and Claude Opus 5, rather than later models named in the source material, including GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 and Gemini 4 Argon. Kingy AI reported Terminal-Bench scores of 57 to 66 for current frontier models in separate evaluations, but those results are not directly comparable with DeepSeek’s test.

It is also not clear from the available information exactly what training changes produced the benchmark gains, how broadly the reported task costs apply, or how 1017 performs across larger and more varied agent workloads. Differences in harnesses, prompts, task selection and usage patterns could affect both scores and costs.

API Use Will Test the Gains

Developers can now access 1017 through DeepSeek’s website, app and API under the existing `deepseek-flash` name, according to the release details. The next useful evidence will come from independent benchmark runs and real deployments that report task success, latency and total cost using clearly described methods.

DeepSeek’s own harness recommendations and support for CodeBuddy and WorkBuddy may guide early use, but the supplied material does not give a date for further model updates or independent evaluations. Users comparing providers will need to check the relevant model versions, pricing periods and test conditions as new results appear.

Key Questions

What is DeepSeek V4.1 Flash 1017?

It is DeepSeek’s Oct. 17 update to V4.1 Flash. The company says it replaces the September version on its website, app and API, with existing API names routed to 1017.

Did DeepSeek raise its API prices?

No price increase is reported. The listed token prices are unchanged, though charges vary by input or output, cache status and whether usage falls in peak or off-peak hours.

Are the reported benchmark improvements independently verified?

Not in the supplied material. The 11 benchmark figures are from DeepSeek’s model card. Kingy AI reported a separate, small pilot, but it does not reproduce the full benchmark table.

Does 1017 outperform the leading models overall?

The available chart does not show that. DeepSeek’s table has 1017 leading four of 11 rows; Claude Opus 5 leads six and GPT-5.6 Sol leads one. The comparison also does not include several newer models named in the source material.

What remains unknown about the update?

DeepSeek has not detailed the training changes behind the gains in the supplied material. Broader independent results, performance across production workloads and how widely the reported per-task savings apply also remain unclear.

Source: ThorstenMeyerAI.com