DeepSeek releases V4.1-Flash-1017 and opens its Search API at $0.75 per 1000 calls

On October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017, an updated Flash model in the V4.1 line. According to the company, it posts higher benchmark results, finishes tasks faster and writes shorter answers than the previous version. The model already runs on the website, in the app and in the API, and the weights are open. At the same time, DeepSeek made its Search and Fetch APIs generally available at $0.75 per 1000 calls per endpoint.

Pareto improvement: quality goes up, cost per task goes down

The release's main claim is that Flash-1017 beats its predecessor on two axes at once: performance and cost. In economics this is called a Pareto improvement: a gain in one dimension is not bought with a loss in another. That matters for agentic workloads. In these tasks, tokens go not to a single answer but to long chains of steps, so shorter outputs directly shrink the bill.

The company shows a DeepSWE v1.1 chart: Pass@1 against the cost of a single task. The measurements were run on the mini-swe-v2 harness. The announcement gives no exact figures, so for now you will have to gauge the scale of the improvement yourself, on your own tasks.

The model was optimized for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. For best results, DeepSeek recommends updating Harness to v0.2.1-rc.1 or newer.

Building projects from scratch with RL and self-feedback

Flash-1017 was trained with reinforcement learning and self-feedback. The goal is end-to-end project development, from games to 3D modeling. DeepSeek showed four examples:

• a 3D model of the Forbidden City;
• a browser racing game;
• a recreation of Portal 2 on the Godot engine;
• Valley of Ages, a 3D scene about evolution from prehistoric times to the present day.

These are demos from the developer, and there is no independent verification yet. Still, the choice of examples shows the direction: the company wants the model to keep a whole project in mind, not just a single function. For 3D scenes and games, that means tying together code, assets, scene structure and logic. The release title says as much: wish-driven programming.

Search and Fetch: a proprietary search engine as a product

The second part of the announcement is commercial. The Search and Fetch APIs are now open to everyone. DeepSeek says they run on its own search engine, which serves hundreds of millions of queries worldwide every day. It is a notable signal: the company is selling not only a model but also the infrastructure an agent uses to reach the internet.

The terms are as follows:

• price: $0.75 per 1000 calls per endpoint, meaning Search and Fetch are billed separately;
• free period: calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users;
• billing start: October 31, 2026, 16:00 UTC;
• documentation: the API Guides section, Search & Fetch.

The mention of Claude Code and Codex is telling. DeepSeek is plugging into other companies' coding agents and competing for the "model plus search" layer where developers already work.

How to try it and what to watch

The model is available on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope, so you can deploy it locally or on your own cluster. If you use DeepSeek Harness, update it to v0.2.1-rc.1 or higher first.

If you are building agents with web access, you have two weeks of free testing for Search and Fetch. In that time, it is worth measuring how many calls each task makes. At $0.75 per 1000 calls per endpoint, an agent that runs dozens of searches and page fetches per task will add a noticeable line to the budget.

It is better to compare Flash-1017 with the previous version by the cost of a solved task, not by the price per token. That is the metric DeepSeek put on its chart, and it is the one that shows whether the promise of shorter outputs and faster completion holds up.