A neat row of citations can make an AI answer look thoroughly checked. Click through, though, and the interesting question is sometimes still waiting: does this page actually say that?

DeepSeek’s latest release gives developers a useful reason to ask. Its 17 October announcement opens access to its own search engine and refreshes Flash. The practical opportunity is better access to source material. Our view is that the important measure is how faithfully an answer uses it.

Scope: This is an analysis of published documentation and vendor claims, checked on 18 October 2026. AI The Thing has not tested these APIs or independently reproduced the model’s results.

Fictional Paper Finch example: an AI answer says commercial use is free and cites a launch post, but current terms require a paid plan. The link works; its evidence does not support the answer.
Illustrative example, not a DeepSeek response. Paper Finch and the documents shown are fictional. The error is relying on an old announcement instead of checking the applicable terms.

Search finds the page. Someone still has to read it.

The Search API returns ranked, structured results. The separate Fetch API returns a page’s title and content as Markdown. That separation matters: locating a promising result and obtaining its contents are different operations.

DeepSeek’s guide describes relevance and authority signals, query operators and extraction failures. Those signals can help choose what to read; they do not certify a statement as true. A page can also be unavailable, blocked or unsuitable for extraction. An application needs to keep that missing evidence visible instead of quietly treating a search snippet as a substitute.

The useful chain is straightforward:

  1. Find locate a relevant source, preferably the original policy or documentation.
  2. Read obtain the passage, including its conditions and exceptions.
  3. Check establish the date, applicable plan and any conflicting update.
  4. Answer make only the claim that the passage supports, or say what remains unknown.

A citation is the start of that inspection. It should not be the decorative full stop at the end of it.

The price of looking something up

Retrieval, before the answer

Search
$0.75 / 1,000 successful requests
Fetch
$0.75 / 1,000 successful requests

Illustrative calculation: 20 qualifying searches + 20 qualifying fetches = $0.03 in retrieval charges.

These dollar figures reproduce DeepSeek’s displayed rate card; they are not AUD conversions. Model generation is charged separately, with peak and off-peak token rates. Additional lookups and retries can change the total. The guide specifies which successful, nonempty outcomes qualify for billing.

Official pricing and billing notes. This is arithmetic, not a bill from a test we ran.

There is also a limited promotion. The launch announcement and pricing page say the specified integrated server-side Search and Fetch calls through DeepSeek Harness, Claude Code or Codex are currently free, with charging starting at midnight Beijing time on 1 November 2026, or 02:00 AEST. This does not make standalone API calls generally free. Recheck the dated rate card before budgeting.

The measure we would use is cost per answer that survives checking. If a cheap response requires five minutes of repair, the interesting part of its price is probably no longer the fraction of a cent spent finding the page. Our inference-cost update explains why the denominator matters.

Flash has changed, even though its API name has not

The October refresh is DeepSeek-V4.1-Flash-1017, served through deepseek-flash. DeepSeek says the calling method and prices are unchanged. A stable alias is convenient for an integration, but an evaluation should record the model revision and date, rather than assuming yesterday’s name guarantees yesterday’s behaviour.

The publisher’s model card reports Terminal-Bench 4.0 rising from 31.2 to 42.3, and DeepSWE v1.1 from 74.2 to 77.0. It identifies Claude Code and Mini-SWE-V2 as the respective harnesses. These are DeepSeek’s reported results, not our measurements or a controlled comparison we conducted.

They are worth investigating for coding workflows. They do not answer our web-evidence question. Completing a repository task and faithfully resolving a contradiction between two sources need different checks. Nor should scores from different benchmark versions be spliced together into a convenient progress chart.

What would convince us?

We would start with a small set of questions whose answers have already been checked against public sources. Include changing prices, version-specific features, superseded policies and questions for which the sources do not supply an answer. Repeat the same tasks, and keep failures rather than selecting the most flattering response.

For each answer, the review would ask:

That is a proposed evaluation, not a test result. A small case study could expose useful failure patterns, but it would not establish which model is best at research in general.

Our recommendation for anyone building this into a workflow: retain the evidence needed to audit consequential answers, expose failed retrieval and let the system say it cannot establish a fact. Apply appropriate privacy and retention controls to those records. More context is useful only when the answer uses it correctly, a distinction also explored in our RAG production analysis and agent reliability review.

DeepSeek has made another search route available. The useful next step is to inspect the journey from result to conclusion. The link should earn its place beside the sentence.

Sources and editorial method

Primary material: DeepSeek’s dated changelog, Search & Fetch guide, rate card, API references linked above and Flash1017 model card. The release is dated 17 October; an exact announcement time was not established. Prices and documentation can change.

AI assisted source research and drafting. Vendor claims are attributed; the Paper Finch scenario is fictional. No model outputs, latency measurements or original API tests are presented.