DeepSeek's V4.1-Flash update tops its own coding chart, and its search engine is now for sale to every developer

DeepSeek on Saturday, Oct. 17, 2026, released DeepSeek-V4.1-Flash-1017, an update to its V4.1-Flash model that it says is stronger, faster and terser, and opened its in-house web Search and Fetch APIs to all developers.

The model bump is the headline, but the quieter move is the search engine. Agents need fresh web results, and DeepSeek just put its own index on the menu at 75 cents per thousand calls, and made it work inside Claude Code and Codex for free for two weeks. That's a Chinese lab shipping plumbing to rivals' coding tools, not just weights. The benchmark table is DeepSeek grading its own homework, and Opus 5 still wins several rows of it. But near-frontier coding scores at cents per task keep putting price pressure on everyone else.

DeepSeek released DeepSeek-V4.1-Flash-1017 on Oct. 17, 2026. The Hugging Face model card calls it a "minor update" to DeepSeek-V4.1-Flash. It is live on DeepSeek's website, app and API. Those lines are DeepSeek's.

DeepSeek says the update delivers stronger benchmark scores, faster task completion and shorter outputs than the previous version — what it calls "a Pareto improvement in performance and cost." Pareto, here, means better on both axes at once: more capability for less spend. Those lines are DeepSeek's.

On DeepSeek's own table, DeepSWE v1.1 is 77.0, up from 74.2 on the previous V4.1-Flash, against Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. Terminal-Bench 4.0 is 42.3, up from 31.2; Opus 5 is still ahead at 51.8. Humanity's Last Exam without tools is 42.6; Opus 5 is 56.3. Apex Bench v1.1 is 72.1 against Opus 5's 65.8. Agents' Last Exam is 35.4 against Opus 5's 28.6. These are company-reported numbers, not independent results. In the same table, Opus 5 still leads on Humanity's Last Exam, Terminal-Bench, ProgramBench and NL2Repo.

On DeepSeek's DeepSWE chart — pass rate versus average cost per task, using a standard software-agent harness — the new model's top setting reaches 77.0% at roughly 20 cents per task. The Claude and GPT models plotted on that chart sit at dollars per task. Those figures are DeepSeek's.

DeepSeek says the model used reinforcement learning with self-feedback to get better at building whole projects end to end. Its demos include a 3D model of the Forbidden City, a web racing game and a Portal 2 recreation in the Godot engine. Reinforcement learning, here, means the model practices and scores its own work. Those lines are DeepSeek's.

The model is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. Weights are on Hugging Face and ModelScope under the MIT license. The Hugging Face repo lists about 763 billion parameters. A parameter is one of the numbers the model learned. MIT license means anyone can use, copy and sell the weights.

New this drop: DeepSeek opened its Search API and Fetch API to everyone. Both sit on DeepSeek's in-house search engine, which it says already serves hundreds of millions of requests a day. Search returns structured web results for a query. Fetch returns a single web page as Markdown, a plain-text format that agents can read. Price: $0.75 per 1,000 calls per endpoint. Those lines are DeepSeek's.

Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users for now. Billing starts Oct. 31, 2026, at 16:00 UTC. Those lines are DeepSeek's.

Bloomberg independently reported the same Saturday that DeepSeek pushed out an update it says is faster and leaner, live on the website, app and developer platform.

In plain terms, DeepSeek said on Saturday that a minor update to V4.1-Flash is live, with MIT-licensed weights and a company table that puts it ahead of its old score on several agent and coding tests, though Opus 5 still wins several rows. It also put its own web search and page-fetch APIs on sale at 75 cents per thousand calls, free inside a few coding tools until the end of October.