Deepseek V4.1 Flash 1017 targets lower-cost coding agents

Deepseek has released Deepseek-V4.1-Flash-1017, an updated Flash model aimed at coding and agent workflows. The company says the model is faster, more concise and cheaper to run per task, which matters for developers building software agents that call a model many times.

The release also includes new Search and Fetch APIs, giving outside developers access to Deepseek’s in-house web search system. For app builders, the pitch is not just a stronger model, but a more complete stack for agents that need to research, code and take multi-step actions.

A Flash model built for end-to-end coding tasks

Deepseek-V4.1-Flash-1017 is available through the company’s website, app and API, according to Deepseek. The company has also published open weights on Hugging Face and ModelScope.

Compared with the previous V4.1-Flash model, Deepseek says the new version improves benchmark scores, completes tasks faster and produces shorter answers. The company describes the update as a better balance between performance and cost.

A key part of the release is reinforcement learning with self-feedback, or RLSF. Deepseek says this post-training approach lets the model use its own confidence signals or self-generated critiques as a reward signal, rather than relying on human evaluators or external reward models.

The company says that method helps with complete software projects, including game programming, 3D modeling and design work. The model can also create compact animated explainer videos, according to Deepseek.

Deepseek says the model was tuned for development environments including Deepseek Harness, Tencent CodeBuddy and WorkBuddy. For Deepseek Harness, the company recommends version v0.2.1-rc.1 or later.

The cost claim centers on shorter outputs

Deepseek’s internal benchmarks put V4.1-Flash-1017 ahead of its predecessor and several larger rival models in coding and agent categories. The company says it outperforms models including Claude Opus 5, GPT 5.6-Sol, Kimi K3 and GLM 5.3 on several of those tests.

In a Deepseek cost chart for the DeepSWE benchmark, V4.1-Flash-1017 reaches 77.0 percent at an average cost of just under $0.20 per task. In a lower setting, it clears 72 percent at about $0.10 per task. The previous V4.1-Flash model scored 74.2 percent and cost about $0.50 per run in the same chart.

The same chart places rival models such as GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 between $2 and more than $10 per task, with none matching Deepseek’s reported accuracy. These are Deepseek’s figures, and The Decoder notes that independent benchmarks have not yet appeared.

For small teams, the practical point is straightforward: agent costs add up quickly because a single workflow may require many model calls. If a model can finish tasks with shorter outputs and fewer expensive runs, the savings can be meaningful for coding assistants, QA agents, internal tools and automated support workflows.

Search and Fetch APIs make the agent stack broader

Deepseek is also opening its web search tools to developers through new Search and Fetch APIs. The company says the services are based on an internal search engine that already handles hundreds of millions of queries per day worldwide.

Each endpoint costs $0.75 per 1,000 calls, according to Deepseek. The company says API users can access the features for free through server-side tools in Deepseek Harness, Claude Code and Codex until October 31, 2026, at 4:00 PM UTC.

That matters because many useful agents need fresh information, not just model reasoning. A coding agent may need documentation. A marketing workflow may need current pages. A research tool may need to retrieve sources before summarizing. Search and fetch tools can reduce the need for developers to wire together a separate retrieval layer.

What builders should verify before switching

The release fits a broader shift toward lower-cost agent infrastructure, where model quality, tool access and per-task pricing matter as much as headline capability. We have seen similar emphasis on parallel agent workflows in Claude’s dynamic workflows update, while Deepseek is focusing this Flash release on efficiency and coding tasks.

Still, teams should treat Deepseek’s numbers as a starting point rather than a final buying decision. The Decoder specifically cautions that benchmark results do not always match developer experience with real code quality.

There is also a policy issue to watch. The Decoder reports that the Chinese government is planning tighter rules for publishing open-weight models, or could even block some releases, with one factor being whether a model ranks among the top 10 most capable systems. That context makes Deepseek’s open-weight release especially relevant alongside broader questions about China’s draft controls for open-weight AI models.

For now, the update is most relevant to developers and small software teams experimenting with agents that build, test or modify code. The main question is whether the model’s reported efficiency holds up in real projects, not just in Deepseek’s own benchmark charts.