Summary: DeepSeek released DeepSeek-V4.1-Flash-1017, a coding-focused update that improves benchmarks while completing tasks faster and producing shorter outputs — what the announcement calls a Pareto improvement in performance and cost. It is live on the website, app, and API, and the weights are on Hugging Face and ModelScope. For builders, the interesting part is the combination: a cheaper-per-task agent model plus newly open Search & Fetch endpoints.
What the update changes
According to DeepSeek’s announcement, V4.1-Flash-1017 beats the previous version on benchmarks while finishing tasks faster and emitting fewer tokens. Since token count is what you actually pay for on both input and output loops, shorter completions compound across an agent session — a multi-step coding run with dozens of tool calls benefits more from tighter outputs than from a headline benchmark bump alone.
The team attributes the end-to-end project gains to reinforcement learning with self-feedback, aimed at whole-project work rather than single-file edits. The examples in the announcement are ambitious for a Flash-tier model: 3D modeling of the Forbidden City, a web racing game, a Portal 2 recreation in Godot, and a 3D scene tracing history from prehistoric to modern times.
Harness and tooling compatibility
The model is optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. If you run DeepSeek Harness, upgrade to v0.2.1-rc.1 or later — older versions may not handle the new model’s behavior well, and the announcement is explicit that the newer harness gives the best experience.
This continues the trajectory from the July update I covered in DeepSeek V4-Flash 0731: Same Architecture, Sharper Agents: the Flash line keeps its architecture but pushes agent reliability with each release.
Search & Fetch pricing, finally
Two endpoints — Search and Fetch, powered by DeepSeek’s in-house engine that reportedly serves hundreds of millions of requests daily — are now open to everyone at $0.75 per 1,000 calls each.
There is a grace window worth planning around: calls made through server-side tools in DeepSeek Harness, Claude Code, or Codex are free for DeepSeek API users until October 31, 2026, at 16:00 UTC, when billing starts. If your agent does heavy retrieval, that is a full year of free lookups before costs kick in. I looked at where retrieval costs actually hide in Free Search APIs for Agents: Where the Real Cost Hides, and a fixed per-call price like this makes budgeting straightforward.
Open weights, same checkpoint
The weights are on Hugging Face and ModelScope. If you self-host, this is the same checkpoint serving the API — no drift between what you prototype against and what you deploy.
Practical takeaway
If you run coding agents on the Flash tier, this update is mostly upside: better task completion, faster runs, and lower output volume at the same API surface. The concrete next steps are to pin DeepSeek Harness to v0.2.1-rc.1 or later, and to log Search & Fetch usage now so you know your real retrieval volume before the October 2026 billing date arrives.