If you run a coding agent on DeepSeek with web search switched on, a new line item appears on your bill starting today. DeepSeek's Search and Fetch APIs — the tools that let an agent search the web and pull pages as Markdown — cost $0.75 per 1,000 successful requests each, and per-request billing began at 00:00 Beijing time on November 1, 2026 (9:00 AM PDT, October 31). Until that moment, calls made through DeepSeek's own harness, Claude Code, or Codex were free. They are not anymore. [DeepSeek pricing page, via RECATOOLS, Oct 17, 2026; AIdapted news roundup, Nov 1, 2026]
This is not a token price change. It is a third billing dimension, on top of DeepSeek's already unusual by-the-clock token pricing (off-peak rates at half the peak price). Tokens are metered per million; search is metered per request. If your cost model only counts tokens, it now misses a line.
The headline numbers
| Item | Price | Unit | What counts |
|---|---|---|---|
| Search API | $0.75 | per 1,000 successful requests | One request, up to 50 results — same price for 1 result or 50 |
| Fetch API | $0.75 | per 1,000 successful requests | One URL fetched as Markdown |
| Failed or empty requests | $0 | — | Only requests that return something are billed |
| Rate limit | 5 requests/second | per tool, per account | Fetch has an additional per-website limit |
The two tools are priced identically and billed separately: a task that searches and then fetches three pages pays for four requests. The Search API is backed by DeepSeek's in-house search engine, returns a short snippet plus a 600–800 character summary per result "intended to be read by a model," and accepts queries up to 128 characters with operators such as site: and lang:.
flowchart TD
A["Oct 17, 2026<br/>DeepSeek publishes Search & Fetch pricing"] --> B["Oct 17 – Oct 31<br/>Agent tool calls free<br/>via Harness, Claude Code, Codex"]
B --> C["Nov 1, 2026, 00:00 Beijing time<br/>Per-request billing starts"]
C --> D["$0.75 per 1,000 successful requests<br/>Search and Fetch, billed separately"]
D --> E["First bill with the new line item<br/>Check it against your own request counts"]
The per-unit math: the request is the unit, not the result
$0.75 per 1,000 requests is $0.00075 per request. The table below assumes a 30-day month and a steady daily task count. It counts Search and Fetch requests together, since both bill at the same rate.
| Profile | Searches + fetches per task | Tasks per day | Requests per month | New monthly cost |
|---|---|---|---|---|
| Light: occasional lookup | 5 | 5 | 750 | $0.56 |
| Everyday: research-assisted coding | 20 | 30 | 18,000 | $13.50 |
| Heavy: research-first agent loop | 50 | 100 | 150,000 | $112.50 |
| Team: 5 heavy agents | 50 | 500 | 750,000 | $562.50 |
Two things follow. First, for light use this is noise — under a dollar a month. For a heavy agent loop it is a real line item: $112.50 a month before a single token is counted, and it scales with task count, not with model choice. Second, because the unit is the request and not the result, the optimization is obvious: ask for up to 50 results per search instead of firing one search per question, and fetch only the pages the agent will actually read. A loop that searches once per fact pays up to 50x more than a loop that batches — same information, different request count.
flowchart LR
L["Light<br/>750 req/mo<br/>$0.56"]
E["Everyday<br/>18,000 req/mo<br/>$13.50"]
H["Heavy<br/>150,000 req/mo<br/>$112.50"]
T["Team of 5 heavy<br/>750,000 req/mo<br/>$562.50"]
What it means for you
If you're a light user (a few searches a day inside an agent): do nothing. This adds cents, and the free tier of your attention is worth more than the saving. Just know the line exists so the first bill carrying it does not surprise you.
If you're a heavy user (a research-first agent loop running dozens of tasks a day): this is the first bill where search stops being free infrastructure and starts being a metered input. Before you change anything, count: pull your October request volume — the free month is your baseline — and multiply by $0.00075. If the number lands near the heavy row above, the fix is batching (50 results per request) and fetch discipline, not switching models: the charge follows the tool call, not the model.
If you run a team: the charge multiplies per agent and per seat, and it lands on top of token spend that already varies by time of day. Add a search-request counter next to your token counter, or the first person to notice the new line will be finance, not engineering. Note the rate limit too: 5 requests per second per tool is a ceiling on parallel agent fleets, not just a price.
Open gaps: DeepSeek has not said whether the per-request rate varies by peak and off-peak hours the way its token rates do — model it as flat $0.75 until the pricing page says otherwise. The per-website Fetch limit is unpublished in detail, so a loop that hammers one domain may hit a wall before it hits a bill. And this change arrives two weeks after a roughly five-hour DeepSeek outage on October 15, for which the company refunded a week of customer spending (October 8–14) as credit — reliability, not just price, belongs in the same review.
Every figure dated, every claim sourced. The AI Bill is an independent practice — no vendor sponsorship, no affiliate links, no paid placement.
Sources: DeepSeek pricing page, quoted by RECATOOLS, Oct 17, 2026 (Search and Fetch APIs $0.75 per 1,000 successful requests each; per-request billing starts 00:00 Beijing time, November 1, 2026; server-side tool calls via DeepSeek Harness, Claude Code, and Codex currently free of charge; 5 requests/second per tool; Fetch per-website limit; 50 results per search; 600–800 character summaries; 128-character queries); AIdapted, "AI News Nov 1, 2026" (billing from November 1, 2026, Beijing time, confirmed against DeepSeek documentation); DeepSeek outage and refund via RECATOOLS, Oct 17, 2026 (five-hour outage October 15; refund of October 8–14 spending as credit).
Next: Book a free 45-minute teardown — screen-share of your last 3 months of AI spend, fixed-fee quote before anything starts.
