Deepseek Launches V4.1-Flash-1017 with Lower Costs and Public Search APIs
Deepseek has released V4.1-Flash-1017, a smaller model outperforming larger rivals in coding and agentic tasks while cutting costs, alongside a new public search API.
Chinese artificial intelligence lab Deepseek has released an upgraded iteration of its Flash model, bringing notable advancements to coding and agentic workflows. Alongside this model release, the organization is opening its proprietary search engine to external developers for the first time.
The newly introduced Deepseek-V4.1-Flash-1017 is currently accessible via the company's official website, mobile application, and API infrastructure. Additionally, Deepseek has published the model weights as open source on platforms like Hugging Face and ModelScope. When measured against its immediate predecessor, the updated system achieves superior benchmark outcomes, executes tasks with greater speed, and generates more concise responses. The organization describes this update as a Pareto improvement balancing efficiency and capability.
To construct comprehensive software applications independently, Deepseek relies on a post-training technique called reinforcement learning with self-feedback (RLSF). This mechanism enables the model to leverage its internal confidence metrics and self-generated critiques as reward signals, entirely eliminating the requirement for human annotators or external reward systems. According to the creators, this methodology enhances end-to-end development capabilities across diverse fields, encompassing game programming, 3D modeling, design tasks, and the animation of compact explainer videos. The system has been specifically optimized for development environments such as Deepseek Harness, Tencent CodeBuddy, and WorkBuddy.
Performance and Efficiency Metrics
In evaluations conducted by Deepseek, the V4.1-Flash-1017 version surpasses its older counterpart across all tested metrics. Furthermore, it outperforms several larger competitive models—including Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3—in multiple categories related to software generation and autonomous agent operations.
Cost efficiency remains a central focus of the update. Data from the DeepSWE benchmark highlights the impact of concise output generation. The V4.1-Flash-1017 model achieves a leading score of 77.0 percent while maintaining an average expense of under $0.20 per task. Operating on a lower resource setting, it still achieves a 72 percent success rate at approximately $0.10 per run. In comparison, the older version scored 74.2 percent at roughly $0.50 per task. Competing options such as GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 range from $2 to over $10 per task on the same chart, without exceeding the accuracy of the compact Deepseek model. Developers can try top AI models cheaply through one API at https://apixoai.online.
Public Search Infrastructure
In addition to the language model update, Deepseek is introducing its internal web search tools, consisting of Search and Fetch APIs, to the broader developer ecosystem. The company states that these services are powered by a custom-built search engine currently processing hundreds of millions of daily global queries. Implementation instructions are available within the official API documentation, priced at $0.75 per 1,000 calls. API users can utilize these features without charge via server-side tools in platforms like Deepseek Harness, Claude Code, and Codex until October 31, 2026.
What it means for developers
For engineers and builders, the arrival of Deepseek-V4.1-Flash-1017 offers a compelling option for automated coding and agentic workflows that demand high performance at a fraction of standard operational costs. The integration of self-feedback reinforcement learning reduces reliance on expensive human-in-the-loop validation for complex software projects. Meanwhile, the exposure of Deepseek's internal search engine via public endpoints provides an alternative mechanism for web-connected applications. However, developers should note potential regulatory shifts, as reports indicate the Chinese government may soon tighten restrictions or restrict the publication of open-weight models that rank among top global performers.
Source: Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
