DeepSeek Now Bills Agent Web Searches. Price Your Tools In

DeepSeek now charges for the web lookups your agent makes. From 00:00 Beijing time on 1 November 2026, the Search and Fetch APIs bill at $0.75 per 1,000 successful requests, and DeepSeek's own documentation says server-side tool calls made through DeepSeek Harness, Claude Code or Codex had been free until that moment (DeepSeek Search and Fetch docs). If you resell an agent, your unit cost has just gained a second meter, and it is not measured in tokens.

That matters because most founders model an agent as a model price per million tokens. Search and Fetch break that model. They are standalone tools with their own price, their own failure modes and their own rate limits.

TL;DR

• Search and Fetch each cost $0.75 per 1,000 successful requests, billed per request.
• Failed calls, empty results and timeouts are free, and count does not change the price.
• Server-side tool calls through Harness, Claude Code and Codex now bill from 1 November 2026.
• A cheap model with heavy searching can cost more than an expensive model with none.
• Log tool calls per task, cap them and price with headroom, or your best customers become your worst margin.

Disclosure: ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 1 Nov 2026; ZeroRaised has no commercial relationship with the companies named.

Agent tool call billing is a pricing model where a provider charges for each executed tool request, such as a web search or a page fetch, rather than only for the tokens the model reads and writes.

What does DeepSeek actually bill now?#

The Search API finds web pages for a query and runs on DeepSeek's in-house search engine. The Fetch API retrieves a single page as Markdown. Both are raw JSON APIs with the same bearer auth, at POST https://api.deepseek.com/v1/tools/search and /v1/tools/fetch (DeepSeek Search and Fetch docs).

Search accepts a count from 1 to 50 results, defaulting to 10, and supports operators such as site:, lang:, after: and before:. That gives you precision without a price penalty, because the request is the billed unit.

The important line is the billing rule: per successful request, not by tokens and not by the number of results. Search costs $0.75 per 1,000 successful requests, and Fetch costs the same (DeepSeek Search and Fetch docs).

Which calls are free?#

Only successful ones that produce something count. A call is billable when it returns HTTP 200 and a result: non-empty results for Search, page content for Fetch. Failed requests, empty results and timeouts are not billed (DeepSeek Search and Fetch docs).

That rule is fair, but it hides a trap: a badly tuned agent that keeps retrying weak queries still pays for every one that returns something.

There are also operational limits worth building around: Search allows 100 QPS and Fetch 25 QPS per account, plus a per-domain limit on Fetch. Exceeding them returns HTTP 429, and HTTP 402 means insufficient balance (DeepSeek Search and Fetch docs).

Why is the token table misleading?#

The model prices look cheap in isolation. deepseek-flash is $0.15 per million cache-miss input tokens off-peak and $0.30 at peak, with output at $0.60 off-peak and $1.20 at peak. deepseek-v4-pro is $0.66 and $1.32 for cache-miss input and $1.98 and $3.96 for output. Cache hits cost $0.003 and $0.006 on flash. Both models carry a 1M context window (DeepSeek models and pricing).

Off-peak is half of peak, and peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays (DeepSeek models and pricing).

Now add the tool calls. As an illustration, 10,000 searches at $0.75 per 1,000 is $7.50, and the same volume of fetches adds another $7.50, before a single token is generated (DeepSeek models and pricing). Token prices do not appear in that sum at all.

A research agent that runs several searches and a handful of fetches per task pays the tool bill on every run. That is why the LLM routing debate is incomplete on its own: routing between models controls one meter and ignores the other.

DeepSeek also "reserves the right to adjust" prices (DeepSeek models and pricing). Today's $0.75 is a current rate, not a commitment.

How should you meter tool calls?#

Treat a tool call as a billable event in your own system, not as an implementation detail.

1. Log every call with a task ID. Record the endpoint, the query, the status code and whether it produced a result.
2. Count calls per task, not per user. One user can trigger a hundred searches through one workflow.
3. Cap calls in the agent loop. Give each task a hard ceiling on searches and fetches, and return a partial answer when it is reached.
4. Use count deliberately. Since the price does not move with the result count, ask for enough results to avoid a second search.
5. Schedule batch work off-peak. If a job does not need to run at once, the half-price token window is free money (DeepSeek models and pricing).
6. Price with headroom. Add the tool cost to the token cost, then add a margin for retries and bad searches.

The guide to pricing an AI product covers the packaging side, and the guide to reducing LLM costs covers the metering habits. The HubSpot spend caps analysis is the customer-facing version: a visible ceiling makes an unpredictable bill sellable.

What this means for bootstrapped founders#

The lesson is blunt: agent cost is tokens plus tool calls, and the second half is easier to miss because it arrives as a small per-request charge.

• Re-run your unit economics. If your price was built on token maths, the margin you calculated is out of date.
• Charge for outcomes, not for unbounded research. A task-based price with a call budget protects you when a user asks the agent to "keep looking".
• Separate search-heavy features from chat. They have different cost shapes and should not share one credit pool by default.
• Watch the free tier. A free trial that permits unlimited searching converts your best demo into your largest bill.
• Keep a second provider in reach. The rate can change, and the sibling analysis of OpenAI's funding squeeze looks at the money pressure on the biggest AI vendor.

If you sell an agent that browses on the customer's behalf, this shift also asks a product question: is the search part of the value, or is it overhead? If it is overhead, hide it inside a flat task price. If it is the value, meter it and show it.

Where ZeroRaised fits#

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have an AI product with a real pricing model, submit it here and state how you meter usage, because clear limits help buyers trust a bill. The newsletter carries new analysis like this, and the pricing page explains the listing options. We make no promise of traffic or results.

Sources#

• DeepSeek: Search and Fetch guide
• DeepSeek: models and pricing