Deepseek Releases V4.1-Flash-1017, Slashing Costs on Agentic Coding Tasks
Deepseek has launched Deepseek-V4.1-Flash-1017, an open-weight model trained with self-feedback that undercuts major rivals on coding benchmarks while reducing per-task costs to under $0.20.
Chinese artificial intelligence company Deepseek has released Deepseek-V4.1-Flash-1017, an updated lightweight model designed to deliver major performance gains in coding and autonomous agent tasks while dramatically lowering inference expenses.
The model is accessible immediately via Deepseek's official website, application, and API. Deepseek has also published the open-weight model on Hugging Face and ModelScope.
Self-Feedback Training and End-to-End Development
Compared to its direct predecessor, Deepseek-V4.1-Flash-1017 runs faster and generates more concise responses. Deepseek frames this release as a Pareto improvement in balancing performance against compute costs.
To handle end-to-end software development from scratch, the model was trained using Reinforcement Learning with Self-Feedback (RLSF). Under this post-training method, the system evaluates itself using internal confidence scores and self-generated critiques rather than relying on human annotators or separate reward models.
According to Deepseek, this setup bolsters full-cycle project generation across diverse use cases, including game programming, 3D modeling, design tasks, and animating compact explainer videos. The model has also been optimized for developer environments, including Tencent CodeBuddy, WorkBuddy, and Deepseek Harness (recommending version v0.2.1-rc.1 or later).
Benchmark Gains and Cost Reductions
In internal benchmarks, Deepseek-V4.1-Flash-1017 outperformed its predecessor as well as larger competing models, including Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3 across multiple coding and agent benchmarks.
The model's shorter, more concise outputs translated directly into cost savings on the DeepSWE benchmark:
- Top Setting: Achieved a 77.0 percent score at an average cost of under $0.20 per task.
- Low Setting: Cleared 72.0 percent accuracy for approximately $0.10 per task.
- Predecessor Comparison: The previous V4.1-Flash scored 74.2 percent at roughly $0.50 per run.
In contrast, rival systems such as GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 registered task costs ranging from $2 to over $10 on the same benchmark chart without matching the smaller model's accuracy. However, independent evaluations have not yet verified these internal metrics, and developer impressions regarding real-world code quality may vary.
In-House Search APIs and Policy Headwinds
Alongside the model launch, Deepseek opened its proprietary web search infrastructure to developers through new Search and Fetch APIs. The underlying search engine already processes hundreds of millions of queries daily.
The endpoints are priced at $0.75 per 1,000 calls. Deepseek API users can access the features at no charge through server-side tools in Deepseek Harness, Claude Code, and Codex until October 31, 2026, at 4:00 PM UTC.
Despite the release, broader uncertainty surrounds open-weight distribution in China. Reports suggest the Chinese government is weighing tighter regulations or potential bans on releasing open-weight models, with scrutiny particularly focused on systems ranked among the top 10 most capable AI models.
