DeepSeek's update aims at cheaper finished tasks. Its new search API raises a custody question

DeepSeek shipped two things on October 17 that point in opposite directions for an enterprise agent team. One lowers the cost of finishing a task. The other makes it easy to send every question your agent asks the web to a new provider.

A model update measured in finished tasks

DeepSeek-V4.1-Flash-1017 is a minor update to V4.1-Flash, with weights on Hugging Face under the MIT license. The model card reports gains on agent benchmarks: Terminal-Bench 4.0 from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, Automation-Bench from 54.8 to 60.5 and DeepSWE v1.1 from 74.2 to 77.0. The same table shows Claude Opus 5 still ahead on Terminal-Bench (51.8) and ProgramBench (37.0). All of these are DeepSeek’s own numbers, with the harness named for some of them (Claude Code for Terminal-Bench, Mini-SWE-V2 for DeepSWE).

The more interesting claim is not the scores. DeepSeek says the update completes tasks faster with shorter outputs, “a Pareto improvement in performance and cost,” and that it is tuned for specific harnesses: DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. For an agent, fewer turns and fewer output tokens per finished task can matter more than the price per token, which for deepseek-flash is $0.15 to $0.30 per million uncached input tokens and $0.60 to $1.20 per million output tokens, depending on whether the call lands in DeepSeek’s off-peak or peak hours. That is the right unit, cost per completed task, and you can only measure it on your own tasks in your own harness.

The alias already moved

DeepSeek’s pricing page says deepseek-flash now points to V4.1-Flash-1017, and that requests to the retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by the same new model. If your application calls DeepSeek through any of those names, its model changed on October 17 without a deployment on your side.

A better model is still a different model. Output length, formatting, tool-call behavior and refusals can all shift, and parsers or cost forecasts tuned to the old behavior break quietly. Treat the update as a model change that needs a regression run, and if you need stability, pin to self-hosted weights, which change only when you change them.

Search and Fetch as metered primitives

The second release opens the search engine behind DeepSeek’s own apps. The Search API takes a query and returns results with a short snippet for interfaces, a longer summary written for models, an optional full page, a relevance score and an authority tier. The Fetch API takes one URL and returns the page as Markdown, subject to the target site’s availability and access policies. Each costs $0.75 per 1,000 successful requests; only calls that return results are billed. Rate limits are 100 queries per second for Search and 25 for Fetch, plus a per-domain limit.

Server-side calls made through DeepSeek Harness, Claude Code or Codex are free for now. Billing starts at 00:00 Beijing time on November 1, 2026, which is 16:00 UTC on October 31.

At that price, search is a rounding error in an agent’s budget: a research task with 30 searches and 20 fetches costs under four cents in tool calls. The expensive parts are elsewhere.

Two costs the price list does not show

The queries are the data. An agent’s search queries carry its task: a supplier being vetted, a molecule, an acquisition target, a customer’s name. Sending them to a search provider creates a new record of what your company is researching, held by that provider. DeepSeek’s search guide says nothing about retention or where requests are processed. For many enterprises that is the deciding fact, before quality or price, and it applies to a hosted search API even when the model itself runs on your own hardware.

The results are untrusted. Fetched pages are written by strangers. An authority tier helps rank sources; it does not make a page safe to act on. Anything an agent fetches should enter its context as data with provenance attached, never with the power to add tools, change scope or trigger an action, the same rule that applies to logs and tool outputs.

What to do this week

If DeepSeek is already in your stack, check which names your code calls and run your evaluation set against the new model now. If you are considering it for coding agents, measure cost and success per task in the harness you will use, at the hours you will run. And before any agent’s searches route to a new provider, including through a coding tool’s server-side search, ask where the queries go, how long they are kept and under which jurisdiction. Free until November 1 is a trial period, not an architecture decision.