DeepSeek V4.1 Flash 1017: Cheaper Agent Work and a Search API You Can Rent
DeepSeek released an open-weights update aimed at agent and coding tasks, claims about $0.20 per task, and opened its search engine to outside developers at $0.75 per 1,000 calls.
DeepSeek has published V4.1 Flash 1017, an updated model tuned for agent work, meaning tasks where the AI plans several steps, uses tools and keeps going without you steering. The weights are open, so developers can download them from Hugging Face and ModelScope. On the same day the company opened its own web search engine to outside developers.
Start with the numbers, which come from DeepSeek itself and have not been independently checked yet. On DeepSWE, a test of real software-repair jobs, the model scores 77.0 percent at its top setting, against 74.2 percent for its predecessor. DeepSeek says it beats several larger rivals, including Claude Opus 5 and GPT 5.6 Sol, in a number of coding and agent categories. The part that stands out is price: just under $0.20 per task at the top setting, and about $0.10 on the low setting, which still scores above 72 percent. The predecessor cost around $0.50 per run. The article puts rivals at roughly $2 to more than $10 per task.
The search side is easy to overlook. DeepSeek says its in-house engine handles hundreds of millions of queries a day. The new Search and Fetch endpoints (one finds pages, one reads a page) cost $0.75 per 1,000 calls. Until October 31 the features are free for DeepSeek API users when used through server-side tools in tools such as Claude Code and Codex.
What’s behind it: an agent is only as useful as what it can look up. Running your own search index is expensive, so selling search by the call lets small developers build “research” agents without that burden. At the same time, DeepSeek keeps competing on cost per finished task rather than on raw score, which is how many people actually judge a model.
A few caveats are worth knowing before you get excited. The benchmarks are the company’s own. The article does not name the licence or give model size, so check the Hugging Face page before building anything commercial. The free search window ends on October 31. And the article notes that China reportedly plans to tighten rules on publishing open-weight models, which could affect future releases.
What this means for you: if you only chat with AI in an app, nothing changes today. If you build small tools or run coding agents, this is a cheap model to test, ideally on a few of your own tasks rather than trusting charts. Everyone else can read it as a trend: capable agent models keep getting cheaper.
Sources
Source: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
Four Gemini 4 Agents Beat Minecraft Together, With No Human Help
Google DeepMind researchers report that four Gemini 4 Argon agents defeated the Ender Dragon in 27 hours, dividing tasks and warning each other along the way.
