DeepSeek v4.1 Flash 1017 beats larger models in agentic tasks while cutting cost (the-decoder.com)

Chinese AI company Deepseek has launched Deepseek-V4.1-Flash-1017, an updated model that outperforms larger rivals in agentic tasks while significantly reducing costs. This new version is available on Hugging Face and ModelScope, and has been trained using reinforcement learning with self-feedback, enabling it to complete coding and software development tasks faster and with shorter outputs. In internal benchmarks, it surpassed competitors like Claude Opus 5 and GPT 5.6-Sol, achieving a top score of 77% on coding tasks at an average cost of under $0.20 per task, demonstrating a substantial efficiency improvement compared to its predecessor. Additionally, Deepseek has introduced its in-house search engine as public APIs, allowing developers to integrate powerful search capabilities into their applications. Each API call costs $0.75 per 1,000 calls, but users can access features for free through certain tools until late 2026. While this release reflects significant advancements for Deepseek, upcoming regulations from the Chinese government on publishing open-weight models could impact future accessibility and development within the AI landscape.