"Wish-based" programming capability upgrade, DeepSeek launches: open self-developed search API, only 5 yuan for a thousand calls

"The big one is coming." A sudden outage of DeepSeek a few days ago led many AI enthusiasts to speculate that the domestic large model leader DeepSeek would soon release a significant update, and indeed it has arrived.

On October 17, reporters from The Paper learned that the domestic large model leader DeepSeek has once again released a version update: DeepSeek V4.1 Flash has now been updated to version V4.1-Flash-1017, with the official web version, APP, and API models all synchronized to the new model.

According to the official introduction, compared to the previous version DeepSeek V4.1 Flash, the new model has made significant improvements in various benchmark tests.

While enhancing performance, it shortens the task runtime and output length, allowing users to complete tasks at a lower cost and faster speed, achieving a stronger cost-performance ratio under multiple Harnesses. (The official term is "performance---cost Pareto improvement.")

The so-called Harness refers to a set of software support frameworks built around large language models, aimed at helping AI Agents manage and execute more complex tasks. Simply put, it can be understood as the complete operating system needed for a large model to truly become an Agent.

It is worth noting that the DeepSeek Harness team also published a statement on the same day: the newly released DeepSeek Harness version 0.2.1 is compatible with the newly released DeepSeek-V4.1-Flash-1017 model. The new model has undergone special optimization for DeepSeek Harness during training, improving task performance while shortening task runtime, reducing output consumption, and token costs.

The first login to use DeepSeek Harness will receive a bonus of 6 yuan, no recharge required, with limited issuance of bonuses, available until exhausted, valid for 7 days after receipt.

In terms of capabilities, DeepSeek V4.1 Flash 1017 enhances the ability to build complete projects from scratch in areas such as game development and 3D modeling through self-feedback-based reinforcement learning (RL) training technology, supporting more "wish-style" programming tasks and project development requests, with a focus on optimization for platforms like DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.

"Wish-style" programming refers to users describing development needs in natural language, allowing AI to complete development autonomously, reducing human intervention.

Palace Museum 3D modeling display effect

Web-based racing game development effect

DeepSeek also announced the opening of its self-developed search API. According to reports, DeepSeek Search is an AI-native general search engine that indexes hundreds of billions of web pages globally, supports multilingual searches, and filters out SEO cheating, GEO manipulation, and low-quality AIGC through strict identification mechanisms, ensuring that authoritative and credible information is presented to users and LLMs.

In terms of pricing, this update is considered a cost-performance king: it is reported that the new version of DeepSeek API has been released, adding support for Search and Fetch interfaces, with related capabilities officially open to the public, both interfaces priced at 5 yuan per thousand calls.

Users can now access the Search/Fetch interfaces through server-side tools when using DeepSeek API in DeepSeek Harness, Claude Code, or Codex, and it is free until November 1.