DeepSeek Search & Fetch Now Bill Per Request: The November 2026 Math

DeepSeek announced Search and Fetch alongside the V4.1-Flash-1017 release on October 17 and ran them free through the launch window. That window closed at midnight Beijing time today. This is what the paid tier actually looks like, straight from the official docs — including the clauses nobody's headline mentions.

What changed on November 1

Two things, at once:

• Direct API calls — anything hitting POST /v1/tools/search or POST /v1/tools/fetch with your own key now bills per successful request.
• Server-side tool calls — per the official pricing page, "when you use the DeepSeek API through DeepSeek Harness, Claude Code, or Codex, server-side tool calls already hit the Search and Fetch APIs. These calls are currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026." If you use DeepSeek as the model backend for Claude Code, every web search the agent fires is now a line item on your DeepSeek balance.

The price is identical for both APIs: $0.75 per 1,000 successful requests, billed against your existing DeepSeek balance (granted balance is deducted first, before your topped-up balance).

The billing rules are unusually friendly

Per-request billing can be a nasty surprise, so the fine print matters. Three clauses in DeepSeek's favor to you:

• Only successful requests bill. A request counts only if it returns HTTP 200 and produces a result — Search must return a non-empty results array, Fetch must return actual page content. Failed requests, empty search results, and timeouts are all free. If you're building RAG over niche queries, you won't pay for dead ends.
• count doesn't affect price. Asking for 10 results per search costs the same as asking for 5. If you've been starving your count parameter to save money, stop — take the fuller result lists.
• Search and Fetch quotas are independent. Heavy fetching doesn't eat your search headroom or vice versa.

There's no free tier mentioned anywhere in the current docs — unlike the model APIs, which have no free quota either, just off-peak discounts. Every successful call draws down your balance.

What it costs you: three honest scenarios

My own arithmetic, not official numbers — your agent's call pattern is the only input that matters:

• Light (50 successful searches/day): ~1,500/month → about $1.13/month. Noise.
• Moderate (500/day): ~15,000/month → about $11.25/month. Noticeable but comparable to a couple of off-peak Flash token runs.
• Heavy agentic (3,000/day): ~90,000/month → about $67.50/month. Now it's a line item worth watching — for context, that's roughly the cost of 45–225M Flash output tokens depending on peak/off-peak timing.

For scale: the deepseek-flash model itself runs $0.15–$0.30 per million input tokens (cache-miss, off-peak vs peak) and $0.60–$1.20 per million output tokens. Search calls are priced per request, not per token, so a single search that retrieves 10 pages costs $0.00075 regardless of how much text comes back. The economics favor Search heavily for anything retrieval-shaped — the expensive part of agentic search was never the retrieval, it's what you feed into the context window afterward.

Baidu is undercutting by 9%

Two days before the billing switch, Baidu launched its own AI search API at ¥4.9 per 1,000 calls — about $0.68 — a price widely reported in Chinese tech press as an explicit counter to DeepSeek's announced $0.75 (see Beijing Business Today's October 30 coverage). A 9% undercut on a commodity API is a classic price-war opening move, and it tells you where this market is going: web search for agents is becoming a metered utility, and Chinese providers are already fighting over the meter rate.

My take: don't switch providers over 7 cents per thousand calls. Switch over result quality and latency. But if you're price-sensitive at high volume, the existence of a cheaper alternative is useful negotiating context — and worth re-checking in a month, because these two are clearly watching each other's price pages.

Rate limits: 100 QPS search, 25 QPS fetch

The official limits are account-level, not per-key: Search allows 100 QPS, Fetch 25 QPS, calculated per account regardless of how many API keys you hold. Fetch also has a separate per-domain QPS limit (undocumented in magnitude) — if you hammer a single website through Fetch, expect HTTP 429 before you exhaust your account quota. Requests rejected with 429 are not successful requests, so they don't bill — but they also don't get your data.

November 1, 2026

Related reading: Search & Fetch launch announcement analysis, DeepSeek V4.1-Flash-1017 release notes, GLM-5.3-Flash vs DeepSeek V4.1-Flash price comparison, and Ling 3.1 Flash 1019 verbosity economics.