Sell a Small Research Agent Service on DeepSeek Search: $0.75 Per 1,000 Calls, Real Costs, No Hype

On November 1, 2026, a free window closed that many coding agent users never noticed was free in the first place. DeepSeek's Search and Fetch APIs each cost $0.75 per 1,000 successful requests, and per-request billing for server-side Search and Fetch calls made through DeepSeek Harness, Claude Code, or Codex started at 00:00 Beijing time on November 1, 2026, which was 16:00 UTC and noon Eastern on October 31. Before that moment, those server-side calls were free of charge. Direct calls to the public Search and Fetch endpoints were already billed at the same $0.75 rate. From November 1, there is no free route left for those calls, only a very cheap one, and the difference between free and cheap is exactly what a seller must price.

That price is good news for a small service, if you respect it. Search at $0.75 per 1,000 successful calls is $0.00075 per call. A research agent that runs 100 searches a day costs $0.075 a day in search calls, before model tokens. A research agent that runs 2,000 searches a day costs $1.50 a day in search calls, before model tokens. The second number is still small, but it is 20 times the first, and an agent that searches on every step, fetches every result, and retries failures without a limit can reach the second number without doing 20 times the useful work. The service in this article is built around that discipline: one bounded research task, a call budget, a fetch rule, and a price that survives the end of free search.

What changed on November 1, and what did not?

DeepSeek opened its Search and Fetch APIs to everyone on October 17, 2026, alongside the DeepSeek V4.1 Flash 1017 release. Search locates webpages. Fetch retrieves the contents of a single page. Under DeepSeek's billing rules, a search is billed when it returns results, and a fetch is billed when it returns page content. Failed requests, empty searches, and timeouts are not billed. That failed request rule is generous, and it matters for agents, because a badly written agent that retries a broken query many times does not pay for the failures, but it does pay for every successful retry that returns results, so retries are still not free in practice.

The November 1 change applies only to server-side tool calls made through three routes: DeepSeek Harness, Claude Code, and Codex, when they use the DeepSeek API. Those calls already hit the Search and Fetch APIs behind the scenes, and DeepSeek had said they were currently free of charge until November 1. That free period is now over. The change does not establish a price change for every user of those coding applications, and it does not change ordinary consumer chat subscriptions. It also does not change model token billing, which continues as before, and search and fetch charges come out of the same prepaid balance as token usage, with any granted bonus balance used first, according to DeepSeek's billing notes as reported.

Model tokens are billed separately from search and fetch calls, and that separation is the most common pricing mistake in this space. A search call price of $0.75 per 1,000 does not include the cost of reading the search results, because the model still pays input tokens for the content it reads. A fetch call price of $0.75 per 1,000 does not include the cost of reading the fetched page, for the same reason. When you quote a client, you must quote three costs: search calls, fetch calls, and model tokens, plus your time. A quote that only mentions $0.75 per 1,000 searches is not a quote, it is a headline, and it will be wrong on the first invoice.

Cost table: Search API, direct call: $0.75 per 1,000 successful requests, billed when a search returns results. Fetch API, direct call: $0.75 per 1,000 successful requests, billed when a fetch returns page content. Search and Fetch through Harness, Claude Code, or Codex: $0.75 per 1,000 successful requests, from November 1, same success rules, after the free window ended at 00:00 Beijing time on November 1. Failed, empty, or timed out calls: $0, no results or no page content returned. Model tokens: billed separately, at the model's own rate, for every token the model reads and generates, including search results and fetched pages. Prepaid balance: search, fetch, and token charges share the same balance, granted bonus balance is used first, according to DeepSeek's billing notes.

What service should you sell first?

Sell one bounded research agent, not a general research department. A good first package is a weekly competitor and pricing watch for one small business: the agent checks a fixed list of competitor pages and public sources once a week, summarises what changed, links to the source for every claim, and flags only changes that matter, such as a price change, a new product, or a closed location. That package is small enough to price on real call counts, specific enough for a client to judge, and safe enough to support, because the sources are public, the output is a summary with links, and a wrong summary can be checked against the linked source in seconds.

The workflow has six steps, and each step has a call budget. First, a client call and a written scope: which competitors, which pages, which changes count, and who reads the report. A watch list of 10 pages is a different service from a watch list of 200 pages, and you should not take the larger scope as your first job. Second, build the source list and test each source once, by hand, before you automate it, so you know which pages block fetching, which pages change layout often, and which pages are not worth fetching at all. Third, set the search rule: search only when a direct fetch is not enough, and cap searches per report, for example 50 searches per weekly report, so a runaway agent cannot turn a weekly report into a daily bill. Fourth, set the fetch rule: fetch only the pages on the approved list, and fetch each page once per report, not once per question, so one report does not pay for the same page ten times. Fifth, add source retention: keep the source URL, the publication date where available, and the fetched text for each claim, so the client can verify the summary without trusting the agent. Sixth, deliver with a handover note: how to add or remove a source, what the report cannot prove, what the monthly call budget is, and what happens when the budget is exceeded.

Do not sell the opposite package as your first offer. Do not sell an unlimited research agent that answers anything, monitors the whole web, or replaces a market research team, because a first agent built on cheap search calls cannot support that promise, and a client who buys it will discover the limits on a live question, not in a demo. Do not sell a service that scrapes paywalled content, private data, or sites that block automated access, because DeepSeek's Search and Fetch APIs are not a licence to ignore a site's terms, and a client who asks you to bypass a paywall is asking for a legal problem, not a research service. Do not sell a price prediction or investment advice agent as your first job, because those domains need professional review and liability cover that a search call budget does not provide. Sell a finished weekly watch for one task, with sources, with a call cap, and with a human check before anything important is acted on.

What does the service really cost, and what should you charge?

Start with the costs you can verify. Search and fetch are $0.75 per 1,000 successful calls each. Model tokens are separate, and you must check the model's own rate on DeepSeek's pricing page at quote time, because DeepSeek uses peak and off-peak schedules for some models, and a rate quoted from memory can be wrong by a factor of two. Your main cost is your time: a first weekly watch will take about 12 to 18 hours, including the client call, source testing, building, a test report, and handover, if you are learning the stack as you go. With a template from the first build, a second similar watch can fall to about 6 to 10 hours, but it will not fall to one hour, because the value is in the source work, the call discipline, and the verification, not in pressing a deploy button.

The price illustration below uses a transparent example, not a forecast, because DeepSeek does not publish a per client hosting price that applies to every job, and your client's usage after the first month will depend on their source count, their report frequency, and their model choice. The illustration prices the first package on time and on a small paid usage buffer, so the client is not surprised when the first invoice arrives.

Package table: Package price, example: $299, one weekly competitor and pricing watch, up to 10 sources, one report format, handover note, one revision round. Your time, first build: 15 hours, example, learning included, second similar build should be faster. Implied rate before costs, on this example: $19.93 an hour, $299 divided by 15 hours, before tax, fees, and client finding time. Search calls per weekly report, example cap: 50 searches, 50 times $0.00075 equals $0.0375 per report in search calls alone. Fetch calls per weekly report, example cap: 10 fetches, one per source, 10 times $0.00075 equals $0.0075 per report in fetch calls alone. Search plus fetch per month, at 4 reports: $0.18 a month, four reports times $0.045, before model tokens, which are billed separately and are the larger cost. Paid usage buffer after setup, example: $20 a month, illustration only, client's ongoing search, fetch, and model token usage, priced from actual usage, not from this guess. What the client pays after the package: Actual usage plus any support retainer you agree in writing. Do not promise free search forever, because the free window ended on November 1.

Read that table honestly. Search and fetch together are $0.045 per weekly report in the example, which is almost nothing, and that is the point. The search price is not the reason to sell this service, and it is not the risk either. The risk is model tokens, because every search result and every fetched page becomes input tokens that the model must read, and a report that fetches 10 long pages and asks the model to summarise all of them can cost far more in tokens than in search calls. The discipline in this service is therefore not just a search cap, it is a reading cap: fetch short, relevant pages, summarise in stages, and do not feed the model a whole website when one page answers the question. A seller who only watches the $0.75 line will be surprised by the token line, and a client who only hears the $0.75 line will feel misled when the token line arrives.

Set three controls before you take a client. First, a spend limit on the DeepSeek account, so a runaway agent cannot spend the client's prepaid balance in one night. Second, a call log that the client can see: searches per report, fetches per report, and the resulting search and fetch cost, separate from token cost, so the client can see that you are staying inside the cap you quoted. Third, a monthly review in the first three months: compare actual calls and tokens against the illustration, adjust the source list or the report frequency if the actuals are higher, and reprice the retainer in writing if the scope grows. Those three controls are what turn a cheap API into a trustworthy service, and they are the reason a client will renew after the first month, when the novelty of cheap search has worn off and the quality of the weekly report is the only thing left.

If you already use Claude Code or Codex on a DeepSeek API key, check your own bill this week, not next month. Your web searches stopped being free at noon Eastern on October 31, and your first November invoice will include search and fetch calls that were free in October. Check your agent's tool call logs for search and fetch requests, count them, multiply by $0.00075, and add the token cost for the content those calls returned. That one check is the fastest way to learn whether your agent is a disciplined researcher or an expensive browser, and it is the same check you will run for every client you take on. Free search is over. Cheap, counted search is the business now.

Key takeaways: DeepSeek Search and Fetch each cost $0.75 per 1,000 successful requests from November 1, 2026, after the free window for server-side calls through DeepSeek Harness, Claude Code, and Codex ended at 00:00 Beijing time on November 1. Failed, empty, and timed out calls are not billed. Successful searches and fetches are billed per request, not per result or per token, and tokens are billed separately. A 2,000 search day costs $1.50 a day in search calls alone, before tokens, while a 100 search day costs $0.075 a day, so call discipline, not the headline rate, decides the bill. A realistic first service is one bounded weekly watch, up to 10 sources, with a search cap, a fetch once per source rule, source links for every claim, and a human check before important actions. The illustration prices one package at $299 for about 15 hours of first build work, with search plus fetch at $0.045 per weekly report in the example, before model tokens, which are the larger cost and must be quoted from DeepSeek's current pricing page. Set a spend limit, a visible call log, and a monthly review in the first three months, and check your own Claude Code or Codex search logs this week, because October searches were free and November searches are not.

Related links: Top Use AI Blog, How to make money with AI, Best AI productivity tools

Affiliate disclosure: No affiliate link is used in this article. Links to DeepSeek and to TopUseAI posts earn no commission.

Sources: DeepSeek API Docs, Models and Pricing, Search and Fetch pricing, $0.75 per 1,000 successful requests each, per-request billing from 00:00 Beijing time on November 1, 2026 for Harness, Claude Code, and Codex server-side calls; DeepSeek Search and Fetch guide, billing and rate limits, failed, empty, and timed out calls not billed; DeepSeek V4.1 Flash 1017 release and Search and Fetch billing announcement, October 17, 2026, as reported October 28 to 31, 2026.