Near-Frontier AI Agents For Less? A Look At DeepSeek V4.1 Flash 1017

🔍 Read the full analysis: Near-Frontier AI Agents For Less? A Look At DeepSeek V4.1 Flash 1017 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 while keeping the model’s listed prices unchanged. The company reports higher scores on all 11 benchmarks it published, with the largest gains on some multi-step agent tasks; those results are vendor-reported and have not been independently reproduced.

The update is served as deepseek-flash, and DeepSeek says older names, including deepseek-v4-flash, now route to the new version. The model’s weights are available under the MIT license. DeepSeek says it tuned the release for its own agent harness and for Tencent’s CodeBuddy and WorkBuddy, but has not detailed what changed during training.

In DeepSeek’s published comparison, Terminal-Bench 4.0 increased from 31.2 to 42.3, while ProgramBench rose from 20.3 to 29.1. The gain on DeepSWE v1.1 was smaller, moving from 74.2 to 77.0. These are benchmark score changes, not evidence of equivalent improvements across every real-world use. DeepSeek also reports that the model produces shorter outputs, which it says can reduce completion time and cost.

The model card lists prices per million tokens of $0.15 for uncached input off-peak and $0.30 during peak hours; output costs $0.60 off-peak and $1.20 at peak. Cached input is cheaper still. Peak pricing applies during specified weekday UTC windows, excluding Chinese public holidays. The model supports a reported context window of up to one million tokens, although real costs depend on usage, caching and output length.

Lower Costs for Agent Work

The commercial pitch is cost per completed task, not a claim that 1017 leads every benchmark. The Decoder’s account of DeepSeek’s DeepSWE chart says the new model reached 77% at just under $0.20 per task, while a lower-effort setting scored above 72% at about $0.10. The same account puts the September model at roughly $0.50 per task for a 74.2% score, and selected frontier models between $2 and more than $10. These figures come from a vendor benchmark and should not be treated as a general price guarantee.

If comparable performance holds in independent testing, lower per-task costs could make long-running coding and automation agents more practical for teams with high usage. But benchmark scores and token prices alone do not settle the question: reliability, latency, integration costs and the amount of human review required also affect the cost of a finished job.

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 DeepSeek’s model costing about a quarter as much. The sample is small, and the perfect results did not distinguish the three models. Kingy also found 1017 was not always faster; one task took about 27 seconds compared with roughly three seconds for Sonnet. The pilot offers an initial outside data point, not a broad validation of DeepSeek’s benchmark claims.

What DeepSeek Compared

DeepSeek’s table compares 1017 with its September release and four selected systems: Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. In that table, 1017 leads four of 11 rows; Opus 5 leads six, and GPT-5.6 Sol leads one. Opus 5 has higher scores on the table’s HLE reasoning measure, ProgramBench and NL2Repo-Bench, among other entries. The comparison therefore shows improvement over the prior Flash release without establishing that 1017 is the strongest model overall.

The selected rivals also do not include newer models named in the source material, such as GPT-6.1 Sol, GPT-6 Astra, Opus 5.5 or Gemini 4 Argon. Kingy AI cited separate Terminal-Bench results of 57 to 66 for current frontier systems, but those evaluations are not a direct comparison with DeepSeek’s run. Differences in harness, setup and evaluation conditions can affect scores.

According to a technical report summarized by Kingy AI, the model uses a sparse mixture-of-experts design and accepts images. The report describes a one-million-token context and a compressed key-value cache. DeepSeek has not publicly explained the training changes behind the October update. The Decoder attributed the gains to post-training with reinforcement learning and model self-feedback, but that account is not a detailed technical disclosure from DeepSeek.

“minor update”

— DeepSeek, in its model update description

Independent Testing Still Needed

Independent reproduction of DeepSeek’s 11 benchmark results is not available in the supplied reporting. The company has not specified what training changes produced the gains, and it is unclear how much shorter outputs affect total costs across different workloads. Benchmark conditions also matter: DeepSeek says Terminal-Bench ran in the Claude Code harness and DeepSWE in Mini-SWE-V2, while scores from other evaluations may use different setups.

The comparison does not cover every current rival, and some headline figures in the table use different test subsets. For example, DeepSeek’s HLE score is for the full set, while GLM-5.3’s listed result is text-only; DeepSeek gives separate text-only scores. The early Kingy AI pilot is small and cannot establish performance across larger or more varied tasks. It is also not clear from the supplied material how the model’s results vary by task complexity, language, or production environment.

Watch for Wider Evaluations

The updated model is already available through DeepSeek’s website, app and API, with the old model aliases routed to 1017. Developers can test the MIT-licensed weights or use the hosted service, while accounting for the listed peak and off-peak prices. The next useful evidence will be independent benchmark runs that publish their harnesses, prompts, task mix and costs alongside the scores.

Further reporting may also clarify the training changes, the model’s performance on newer rival systems and whether lower benchmark costs persist in real deployments. Until those details emerge, the confirmed development is an available model update with unchanged list pricing; claims about broader performance and savings remain subject to testing.

Key Questions

What did DeepSeek release?

DeepSeek released V4.1 Flash 1017 on October 17, 2026. It replaced the September Flash version on the company’s website, app and API.

Did DeepSeek change the model’s prices?

DeepSeek’s listed prices remained unchanged with the update. The price per million tokens varies by input caching and whether usage falls in the company’s peak or off-peak hours.

Are the reported benchmark gains independently confirmed?

No independent reproduction of DeepSeek’s 11 published benchmark results is described in the supplied reporting. Kingy AI reported a small pilot, but it does not verify the full benchmark table.

Does 1017 outperform every model in DeepSeek’s chart?

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

Why is DeepSeek emphasizing cost per task?

The company’s case is that a lower cost for completing agent tasks could make repeated coding and automation work more affordable. The figures are benchmark-specific; actual costs and performance can vary by workload and deployment.

Source: ThorstenMeyerAI.com