DeepSeek opens its search API: pricing, Flash 1017 and what to test

What changed on October 17?

DeepSeek’s dated changelog confirms two developments: public access to its search engine through an API, and a new Flash checkpoint across its web app, mobile app and API. It says the model calling method and model pricing are unchanged.

For existing integrations, an unchanged model name can conceal a changed underlying model. Promptwire’s recommendation is to rerun a small set of saved tasks before assuming previous outputs, latency or costs will remain representative.

Public record: DeepSeek — Changelog, October 17, 2026

Search finds pages; Fetch reads one page

The official guide describes Search as a structured web-results API and Fetch as a way to retrieve one page as Markdown. Both use a DeepSeek API key. Search accepts a query, optional result count and operators such as site restrictions; Fetch takes a page URL.

The documented endpoints are POST /v1/tools/search and POST /v1/tools/fetch. Each API is limited to five queries per second per account, independently. Fetch also faces per-domain limits and the target website’s access policies.

An example workflow: a product researcher searches for a vendor announcement, retrieves the original page, then asks a model to summarise it with a source link. That workflow is our illustration of how the pieces fit together, not a product test.

Public record: DeepSeek — Search & Fetch guide and billing

What does the search API cost?

The guide lists Search and Fetch at $0.75 each per 1,000 successful requests. A search needs non-empty results to be billable; a fetch needs returned page content. Failed requests and timeouts are not billed. Model-token charges remain separate.

For illustration, 1,000 successful searches followed by 1,000 successful fetches would cost $1.50 in retrieval fees, before model usage. Budget around completed research tasks: one answer may require several searches, page reads and model calls. This arithmetic is a planning example, not an observed usage bill.

Check the current guide for temporary integration-specific fee waivers before estimating a production bill. An introductory concession should not be treated as the long-term unit price.

Public record: DeepSeek — Search & Fetch guide and billing

What does Flash 1017 add?

DeepSeek’s model card calls 1017 a minor update and publishes MIT-licensed weights. Its reported DeepSWE v1.1 score rises from 74.2 to 77.0, while Terminal-Bench 4.0 rises from 31.2 to 42.3. These are company-reported results, not Promptwire measurements.

Our assessment: those results justify an evaluation, rather than an automatic switch. A benchmark score cannot establish whether an agent will follow your requirements, locate the right evidence or produce changes that a reviewer accepts.

Public record: DeepSeek — Flash 1017 model card and license

A useful first test for a small team

Pick ten recent questions whose answers you already know, including a changed price, an outdated support page and an announcement with regional restrictions. Keep the expected answer and original source beside each question. Give the same questions to your current workflow and the new setup.

Record whether the system selects the original source, distinguishes publication dates from update dates and supports each important statement with the retrieved text. Include a question with no reliable answer: a useful research assistant should identify that gap.

Track retrieval calls, model usage, time to a checked answer and the minutes spent correcting mistakes. Review the actual cited passage, rather than accepting a plausible-looking link. Expand the trial only if the combined workflow improves your work. Promptwire has not run this evaluation or independently tested the APIs.