DeepSeek has released a new version of V4.1 Flash. According to the Chinese AI lab, the update, named 1017, performs better on benchmarks, completes tasks faster, and produces shorter output. In addition, DeepSeek is making its Search & Fetch APIs available to all developers.
DeepSeek-V4.1-Flash-1017 is available immediately via DeepSeek’s website, app, and API. The company praises the model for better performance at a lower cost per task. To assess this, DeepSeek examines the ratio of the Pass@1 score on DeepSWE v1.1 to the cost per task, measured using the mini-swe-v2-harness.
On DeepSWE v1.1, the 1017 update achieves 77.0 percent, compared to 74.0 percent for Claude Opus 5 and 73.0 percent for GPT-5.6 Sol. However, these results come partly from DeepSeek itself; independent validation is still pending.
DeepSeek released the first version of V4.1 Flash in September. At that time, DeepSeek claimed the model outperformed its flagship V4-Pro, while the maximum output cost dropped by about 70 percent to $1.20 per million tokens.
Complete projects from scratch
What has changed technically? DeepSeek uses reinforcement learning with self-feedback to improve end-to-end project development. This should enable the model to build complete projects, ranging from games to 3D models. As examples, the company showcases a 3D model of the Forbidden City in Beijing, a browser-based racing game, and a recreation of Portal 2 built with the Godot engine.
DeepSeek claims this enables “wish-driven programming,” in which a user describes the end result in natural language and an AI agent then carries out the entire development process.
The model is optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. For the best experience, DeepSeek recommends upgrading to DeepSeek Harness v0.2.1-rc.1 or later.
Search & fetch for everyone
In addition to the model, DeepSeek is opening its Search & Fetch APIs to all users. These run on the company’s proprietary search engine, which processes hundreds of millions of requests daily. Each endpoint costs $0.75 per 1,000 calls.
For now, calls made via server-side tools in DeepSeek Harness, Claude Code, and Codex are free for DeepSeek API users. This offer ends on October 31 at 4:00 p.m. UTC.