DeepSeek's Free Web Search Ends on 1 November

DeepSeek's server-side Search and Fetch calls stop being free at 00:00 Beijing time on 1 November 2026, which is 16:00 UTC on 31 October and 21:30 IST the same evening. From that moment, every successful search or page fetch an agent performs through DeepSeek Harness, Claude Code or Codex is billed at $0.75 per 1,000 requests. The standalone Search and Fetch APIs were always billed; what ends is the carve-out for tool calls an agent makes on your behalf.¹

This is a small number attached to a large change in behaviour. A coding agent that searches the web ten times per task and runs two hundred tasks a month makes two thousand calls. At $0.75 per thousand, that is $1.50, which is not the point. The point is that a free tier quietly ended and that the bill arrives through a channel most people are not watching.²

The billing that starts on 1 November

Prices measured on 27 October 2026. Vendor plan terms, allowances and prices change without notice, so treat every figure on this page as a dated snapshot rather than a permanent fact. Where a number here differs from the vendor's own pricing page, the vendor's page is correct and the figure above is simply older than it.

Two details in that table are worth reading twice. The first is successful requests. DeepSeek bills per successful request, so a search that errors out is not charged. That is a more forgiving rule than per-call billing and it is stated plainly rather than buried.

The second is that Search and Fetch are charged per request, not per token. A search that returns fifty results costs the same as one that returns three. If your agent over-fetches to be safe, that habit is now metered per call rather than absorbed.

Who actually pays

The note on DeepSeek's pricing page is specific about the path that changes. When you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit the Search and Fetch APIs. Those calls are currently free of charge, and per-request billing starts at midnight Beijing on 1 November.

That describes a growing and slightly invisible population: developers who point an existing agent harness at DeepSeek because the model is cheap, and let the harness do its own searching. The tokens were always the visible cost. The searching was not, because it was free.

If you call the Search and Fetch endpoints directly from your own code, nothing changes for you. Those were billed per request already, at the same rate. The change is confined to calls made on your behalf by a harness.

What the tokens cost, for scale

Search is the new line item, but tokens remain the bulk of any real bill, and DeepSeek's price structure has a wrinkle worth understanding before you estimate anything.

Off-peak rates are exactly half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour is off-peak, including weekends and the holidays in full.

For a developer in India that works out to the early morning and lunchtime slots costing double: 06:30 to 09:30 and 11:30 to 15:30 IST. For a developer in California, peak runs from 18:00 to 21:00 and 23:00 to 03:00 Pacific, which is to say the evening and the middle of the night. Scheduled batch jobs moved a few hours later get the off-peak rate for the same work.

Cache hits are the other lever and they are an order of magnitude apart: $0.003 against $0.15 on Flash, a factor of fifty. Anything that reuses a stable prefix — a system prompt, a repository context, a long instruction file — should be structured so the cache can hit it.

What stays free

The chat app. DeepSeek's consumer product publishes no message cap and has no subscription tier to upsell you into, which is the reason it appears in this directory at all. Nothing in the Search and Fetch change touches that.

Worth stating because the two get conflated in coverage of any DeepSeek pricing story. The app is a free product with no paid tier. The API is a metered product with a prepaid balance. This change moves a line item inside the second one.

How to estimate your exposure

Count the searches, not the tokens. If you run an agent through a harness, look at how many tool calls a typical task makes and multiply by how many tasks you run in a month. The result is almost certainly a small number of dollars, and that is the useful finding: the cost is not the risk, the unpredictability is.

The risk is an agent loop. An agent that fails to find what it wants and retries a search is now spending money on each attempt, and a pathological loop can make hundreds of calls before a human notices. There is no cap on Search and Fetch, because there is no published limit to hit — they are billed per successful request with a concurrency limit of 2,500 on Flash and 500 on Pro.

The practical mitigation is the one that was already good practice: bound the number of tool calls per task in the harness. A limit of ten searches per task turns an unbounded risk into a fixed one.

Balance and deduction order

Search and Fetch fees are deducted from your topped-up balance or granted balance, with the granted balance used first when both exist. If you have promotional credit sitting on the account, it will be consumed by search fees before your own money is touched.

That is worth knowing if you have been treating granted balance as a buffer for model calls. It will now be spent on the cheapest item on the bill first, which is not how most people assume a credit balance drains.

Dates that mattered before this one

Pricing changes keep arriving with a date attached rather than a headline. ChatGPT's Pro 200 allowance drops on 29 October 2026 at an unchanged price, which is the same pattern in a different shape: not a new line item, but less of something you already pay for.

Cursor's repricing into The X Subscription was the loud version, with new tier names and a visible price change. The DeepSeek change is the quiet version: a free carve-out closes and a line appears on a bill most people read once a month.

Frequently asked questions

When does DeepSeek Search billing start?
00:00 Beijing time on 1 November 2026, which is 16:00 UTC and 21:30 IST on 31 October.

What does a search cost?
$0.75 per 1,000 successful requests, for both Search and Fetch. Failed requests are not charged.

Does this affect the DeepSeek chat app?
No. The app remains free with no subscription tier. The change applies to server-side tool calls made through the API via DeepSeek Harness, Claude Code or Codex.

I call the Search API directly. Does anything change?
No. Direct calls were already billed per request at the same rate. What ends is the free carve-out for calls an agent makes on your behalf.

Are model tokens billed differently at peak times?
Yes. Off-peak rates are exactly half of peak. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays.

Is there a cap on search spending?
No. Search and Fetch are billed per successful request with no published spending cap. Bound the tool calls per task in your harness instead.

Search and Fetch rates, the 1 November billing start, model prices and peak-hour definitions verified against DeepSeek's own API documentation on 27 October 2026. Successful-request billing and the granted-balance deduction order are as stated on that page.

Sources

1. DeepSeek API pricing: api-docs.deepseek.com/quick_start/pricing. Used for: Search and Fetch rates, the 1 November billing start and model prices.
2. DeepSeek Search and Fetch guide: api-docs.deepseek.com/guides/search. Used for: Tool-call behaviour and the free carve-out for harnesses.

All sources are the vendor's own plan or documentation pages, checked on 27 October 2026. Plan terms change without notice, so confirm current limits before you rely on them.