DeepSeek V4.1 Flash 1017: A Lower-Cost Route To Capable AI Agents

🔍 Read the full analysis: DeepSeek V4.1 Flash 1017: A Lower-Cost Route To Capable AI Agents on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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 V4.1 Flash 1017 on October 17, replacing its September Flash model across the company’s website, app and API while keeping token prices unchanged. The update matters to developers building AI agents because DeepSeek reports higher scores on all 11 benchmarks it published and says shorter outputs reduce the cost and time of completing tasks; those results have not been broadly independently reproduced.

The company describes 1017 as a minor update to V4.1 Flash, first released on September 10. Calls to deepseek-flash now route to 1017 automatically, as do the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp. The weights are available under the MIT license on Hugging Face and ModelScope.

DeepSeek’s model card reports gains on each of its 11 benchmarks, but the size varies. Terminal-Bench 4.0 rose from 31.2 to 42.3, and ProgramBench from 20.3 to 29.1. DeepSWE v1.1, a coding benchmark, increased by 2.8 points, from 74.2 to 77.0. These are benchmark scores reported by DeepSeek, not independently verified findings.

DeepSeek also lists prices of $0.15 per million uncached input tokens off-peak and $0.60 per million output tokens off-peak. Peak rates are twice those amounts; cached input is cheaper. The company says outputs are shorter, which can lower the bill for completed tasks. The precise savings for a particular project will depend on its prompts, outputs and usage pattern.

Lower Bills for Agent Runs

Agent systems can make repeated model calls while planning, using tools and checking results. For developers running many such tasks, cost per completed task may matter more than a single benchmark score. DeepSeek’s reported pricing and task-cost figures make 1017 a potential lower-cost option for coding and other multi-step work.

The evidence is promising but limited. The Decoder reported DeepSeek’s chart showing 1017 at 77.0% on DeepSWE for just under $0.20 per task; a low-effort setting scored above 72% at about $0.10. These figures come from DeepSeek’s evaluation. In a separate early pilot, Kingy AI reported that 1017 passed 14 of 14 runs across seven small tasks, at roughly a quarter of the cost of GPT-6.1 Sol and Claude Sonnet 5.5. That pilot is small and does not establish performance across broader workloads.

A Small Model With Long Context

DeepSeek presents Flash as a sparse Mixture of Experts model with image input and a one-million-token context window. Its technical report, as summarized by Kingy AI, describes a Causal Encoder-Decoder design with 40 layers and 384 routed experts, six active per token alongside a shared expert. The report describes a 552-billion-parameter backbone, with about 8 billion active during prefill and 16 billion during decoding. Hugging Face lists 763 billion using a different counting method.

The technical report focuses on KV cache compression, the memory used to keep track of prior tokens. It reports about 890 bytes per token, or less than 1 GB for a full million-token context. DeepSeek’s hosted API supports up to 384,000 output tokens. The company has not specified what changed in 1017’s training; The Decoder attributes the gains to post-training with reinforcement learning and self-feedback, rather than human raters or a separate reward model.

DeepSeek recommends its own DeepSeek Harness version 0.2.1-rc.1 or later and says it tuned 1017 for Tencent’s CodeBuddy and WorkBuddy. Benchmark scores also depend on the testing setup: Terminal-Bench used the Claude Code harness, while DeepSWE ran in Mini-SWE-V2.

“minor update”

— DeepSeek, describing the October release

Evidence Beyond DeepSeek’s Chart

Independent validation remains limited. DeepSeek produced the benchmark table, and the early Kingy AI pilot tested only seven small tasks. Its three tested models all passed every run, so the results cannot rank them. The pilot also found that 1017 was not always the fastest: one task took about 27 seconds, compared with about three seconds for Claude Sonnet 5.5.

DeepSeek’s comparisons also omit newer competitors. Its table includes GPT-5.6 Sol and Claude Opus 5, rather than later models 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 models, but those evaluations are not directly comparable. It remains unclear how 1017 performs across larger independent tests, different agent frameworks and real production workloads.

Testing in Real Agent Workflows

Developers can now try 1017 through DeepSeek’s website, app or API, with existing deepseek-flash calls receiving the update automatically. The next useful evidence will come from independent evaluations that publish their prompts, harnesses, task mix and cost calculations, alongside tests on current competing models.

Teams considering the switch can compare completion rates, latency and total cost on their own tasks. DeepSeek’s published results suggest potential savings, but they do not yet establish that 1017 will be cheaper or more capable for every workload.

Key Questions

When did DeepSeek release V4.1 Flash 1017?

DeepSeek released it on October 17, 2026, as an update to the V4.1 Flash model launched on September 10.

Did DeepSeek change the API price?

DeepSeek says the update replaced the previous version at unchanged token prices. Rates vary by input or output, cache status and whether the request falls in peak hours.

Are the benchmark gains independently confirmed?

The 11 benchmark gains in the model card are DeepSeek-reported. An early Kingy AI pilot offers limited outside evidence, but it covered only seven small tasks run twice.

What does DeepSeek say improved most?

In DeepSeek’s table, Terminal-Bench 4.0 increased by 11.1 points, from 31.2 to 42.3, and ProgramBench rose 8.8 points. DeepSWE v1.1 rose 2.8 points.

Does 1017 lead every model in DeepSeek’s comparison?

No. It led four of 11 rows in DeepSeek’s table. Claude Opus 5 led six, and GPT-5.6 Sol led one; the table does not include several newer models.

Source: ThorstenMeyerAI.com