DeepSeek released V4.1-Flash-1017 on Saturday, October 17 — an updated checkpoint of its Flash model that, in the company’s own testing, scores 77.0 on the DeepSWE coding benchmark at an average cost of just under $0.20 per task. Why it matters: the update improves agentic coding scores while cutting per-task cost sharply, and DeepSeek paired it with a first — opening its in-house web search engine to outside developers.
October 17, 2026
What happened
The new checkpoint is live now on DeepSeek’s website, in its app, and through its API under the existing deepseek-flash model name. DeepSeek also published the model weights on Hugging Face and ModelScope. The company describes the release as a Pareto improvement over the September V4.1-Flash: better benchmark scores, faster task completion, and shorter outputs, which is where much of the cost saving comes from.
For end-to-end project work — building complete software projects from scratch, from games to 3D modelling — the checkpoint was post-trained with reinforcement learning with self-feedback, a method that uses the model’s own confidence scores and self-generated critiques as the reward signal instead of human evaluators or a separate reward model. DeepSeek says the model is tuned for its own Harness tool, as well as Tencent CodeBuddy and WorkBuddy.
The numbers, with the caveat attached
All of the figures below are DeepSeek’s own; independent benchmarks have not been published yet. In the company’s table, V4.1-Flash-1017 beats its predecessor on all 11 reported benchmarks, by 2.0 to 11.1 points. On DeepSWE v1.1 it scores 77.0, up from 74.2 for the previous Flash and ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 in the same table. Terminal-Bench 4.0 jumped to 42.3 from 31.2. At a lower reasoning setting, the model still cleared 72% on DeepSWE at roughly $0.10 per task, while the predecessor cost around $0.50 per run.
The lead is not across the board. In DeepSeek’s own table, Claude Opus 5 still leads on Terminal-Bench 4.0, ProgramBench and NL2Repo, and on raw reasoning without tools — 56.3 to 42.6 on HLE — the gap remains wide. That pattern matters: this is a coding-and-agents checkpoint, not a new frontier reasoning model.
A search engine becomes a product
Alongside the model, DeepSeek opened its Search and Fetch APIs to every developer for the first time. Search returns ranked web results for a query; Fetch returns a single page as Markdown. DeepSeek says the underlying in-house engine already handles hundreds of millions of requests a day. Each endpoint costs $0.75 per 1,000 successful requests — empty results and failures are free — with a limit of 5 queries per second per account. Calls made server-side through DeepSeek Harness, Claude Code or Codex are free through the end of October.
Why it matters
For anyone paying for agentic coding by the task, the economics are the story. A company-reported 77.0 on a real-world coding benchmark at under $0.20 a task — against rival figures in DeepSeek’s chart running from $2 to over $10 — puts fresh pressure on the Flash-tier pricing race, provided independent testing confirms even part of the claim. The search launch may matter just as much in the long run: agent builders can now get the model and a first-party live web layer behind one API key, instead of bolting on a separate search provider. As always with company-reported benchmarks, test it on your own tasks before switching anything in production — that is the whole point of everything AI, tested.
Frequently asked questions
What is DeepSeek V4.1-Flash-1017?
It is an updated checkpoint of DeepSeek’s V4.1-Flash model, released October 17, 2026. It is available on DeepSeek’s website, app and API as deepseek-flash, with weights published on Hugging Face and ModelScope.
How much does V4.1-Flash-1017 cost to use?
In DeepSeek’s DeepSWE cost chart, the model averaged just under $0.20 per task at its top setting and about $0.10 in its low setting. API token pricing and exact costs vary by setting and time of day.
Did DeepSeek beat Claude Opus 5?
On some coding and agent benchmarks in DeepSeek’s own table, yes — including DeepSWE v1.1, 77.0 to 74.0. Opus 5 still leads in that same table on Terminal-Bench 4.0, ProgramBench, NL2Repo and HLE reasoning. Independent benchmarks have not been published yet.
What are DeepSeek’s Search and Fetch APIs?
They are developer endpoints built on DeepSeek’s in-house search engine, opened to all developers alongside this release. Search returns ranked web results and Fetch returns a page as Markdown; each costs $0.75 per 1,000 successful requests.
Sources: The Decoder, MarkTechPost, Bloomberg via Savedelete, AI Weekly