How DeepSeek V4.1 Flash 1017 Compares On Price For Near-Frontier Agents

🔍 Read the full analysis: How DeepSeek V4.1 Flash 1017 Compares On Price For Near-Frontier Agents 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 the September version across its website, app and API without changing the listed token prices. The company reports gains on all 11 benchmarks it published, with larger improvements on some multi-step agent tasks; the results remain vendor-reported and have not been independently reproduced at scale.

The update is a “minor update” to the V4.1 Flash model released September 10, according to DeepSeek. Calls to deepseek-flash now reach 1017 automatically. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are also routed to the updated model. DeepSeek says the model is tuned for its own DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy.

In DeepSeek’s model-card table, Terminal-Bench 4.0 rises from 31.2 to 42.3, while ProgramBench increases from 20.3 to 29.1. The change on DeepSWE v1.1 is smaller, from 74.2 to 77.0. DeepSeek also says its newer model produces shorter outputs, which can lower task costs and time to completion. These measurements are reported by DeepSeek; the source material does not provide independent replications of the full benchmark set.

The published price list is unchanged: input without a cache hit costs $0.15 per million tokens off-peak and $0.30 at peak; cached input costs $0.003 and $0.006; output costs $0.60 and $1.20. Peak pricing applies during two weekday windows, 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. DeepSeek’s stated example of one million uncached input tokens and 200,000 output tokens costs $0.27 off-peak.

Lower Costs for Long Agent Tasks

The update’s clearest commercial case is the cost of completing agent work, rather than a claim to lead every benchmark. The Decoder reported that DeepSeek’s DeepSWE chart puts 1017 at 77.0% for just under $0.20 per task; a low-effort setting reached more than 72% at about $0.10. On that chart, the September model cost about $0.50 per task for a 74.2% score, while frontier models ranged from roughly $2 to more than $10. Those are figures reported from DeepSeek’s evaluation, and the source does not establish that every model was run under identical conditions.

A Kingy AI pilot offers a limited outside check. Across seven small tasks run twice, 1017, GPT-6.1 Sol and Claude Sonnet 5.5 each passed all 14 runs. Kingy AI reported that 1017 cost roughly a quarter as much as the other two in the pilot. The sample is too small to establish broad reliability or rank the models; Kingy AI also said 1017 was slower on one task, taking about 27 seconds compared with roughly three seconds for Sonnet.

For teams running many coding or workflow tasks, even a modest reduction in cost per successful run could affect which model they use by default. But the practical savings depend on the workload, prompt and output lengths, cache use, peak pricing, and whether the model completes the task correctly. The published figures provide a reason to evaluate 1017 on specific tasks, not a guaranteed saving across deployments.

What DeepSeek’s Chart Covers

DeepSeek’s table compares 1017 with its September V4.1 Flash and four selected rivals: Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. In that table, 1017 leads 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. DeepSeek’s HLE result uses the full set, while its text-only score is 44.0 for 1017 and 39.1 for September. The GLM-5.3 figure shown is text-only, so those HLE entries are not directly comparable.

The chart also does not include newer named models such as GPT-6.1 Sol, GPT-6 Astra, Claude Opus 5.5 or Gemini 4 Argon. Kingy AI has cited Terminal-Bench results between 57 and 66 for current frontier models, but those come from separate evaluations and are not a direct comparison with DeepSeek’s table. DeepSeek’s results therefore show how 1017 performed against the models and setups it selected, not its rank against the entire current field.

The model is distributed with open weights under the MIT license, according to DeepSeek and its listings on Hugging Face and ModelScope. The technical description identifies a sparse mixture-of-experts architecture with image input and a one-million-token context. The source material gives differing total parameter counts depending on how they are counted: 552 billion in the technical description and 763 billion on Hugging Face. DeepSeek has not explained in the provided material what changed in training for the October update.

““minor update””

— DeepSeek

Independent Results Still Limited

DeepSeek’s 11 benchmark results have not been independently reproduced in the material available for this report. The methodology behind the reported cost-per-task chart, including how task success, model settings and evaluation costs were handled, is not fully established here. The seven-task Kingy AI pilot is too small to show how the model performs across larger or more varied workloads.

It is also unclear what training changes produced the benchmark gains. The Decoder reported that the improvement came from post-training with reinforcement learning and self-feedback, but DeepSeek did not specify training changes in the supplied material. The source also mentions an announcement of two developer search tools but cuts off before describing their functions, availability or pricing. No further details can be confirmed from that account.

Testing on Real Workloads

Developers can access 1017 through DeepSeek’s website, app and API, where the named Flash aliases now route to the updated version. The company’s recommended harness is DeepSeek Harness version 0.2.1-rc.1 or later. The next useful evidence will come from independent evaluations that publish their task sets, settings, success criteria, timing and cost calculations, alongside tests on real workloads.

For teams considering adoption, the relevant comparison is the cost and reliability of completing their own tasks at the required quality level. Further reporting may also clarify the training changes behind the release and provide the missing details on DeepSeek’s developer search tools.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released it on October 17, 2026. The company says it replaced the September version on its website, app and API.

Did DeepSeek raise its API prices?

No price change is reported. DeepSeek’s listed rates vary by token type, cache status and peak or off-peak period; off-peak output costs $0.60 per million tokens.

Has an independent evaluation confirmed DeepSeek’s benchmark gains?

Not at scale in the source material. DeepSeek’s benchmark table is company-reported. Kingy AI described a pilot of seven tasks run twice, but that sample is too small to confirm general performance.

Does 1017 lead DeepSeek’s published comparison?

It leads four of the 11 rows in DeepSeek’s table. Claude Opus 5 leads six and GPT-5.6 Sol leads one; the chart does not include several newer models named in the source material.

What remains unknown about the update?

DeepSeek has not specified what changed in training in the supplied material. Broader independent results, comparable evaluations against newer models and more detail about the announced search tools are also not established.

Source: ThorstenMeyerAI.com