On October 17, 2026, alongside the V4.1-Flash-1017 model update, DeepSeek quietly opened two infrastructure endpoints to everyone: Search and Fetch, both powered by its in-house search engine. The announcement was easy to miss because it arrived under a model-release headline, and because the endpoints were, for most users, free.
That free period has a precise end date. Server-side tool calls that trigger Search and Fetch inside DeepSeek Harness, Claude Code, and Codex are free for DeepSeek API users only until October 31, 2026, at 16:00 UTC — midnight on November 1 in Beijing. After that, each endpoint costs $0.75 per 1,000 calls.
For a chatbot, that is a rounding error. For an agent that searches every few steps, it is a new operating cost that most budgets do not have a line for.
What changed
DeepSeek built its own search engine and says it serves hundreds of millions of requests daily. Opening it as an API turns a proprietary capability into a metered product. The pricing is per call, not per token: $0.75 for 1,000 Search calls and the same for 1,000 Fetch calls. Reports of the CNY list price put it at roughly 5 CNY per 1,000 successful calls.
Three properties distinguish this from ordinary model pricing:
It is per action, not per size. A Search call costs the same whether it returns three results or twenty. A Fetch call costs the same whether the page is short or long. That rewards precise queries and penalizes broad, exploratory ones.
It compounds with the agent loop. Token billing scales with context length; retrieval billing scales with the number of steps. A long-horizon agent that verifies facts at every stage pays both meters, and the retrieval meter grows linearly with the step count rather than with the answer length.
It is billed at the infrastructure layer. Because the calls originate as server-side tools, the cost may not appear in the same place as model tokens. Teams that track only token spend will discover the change on the invoice, not in the dashboard.
The arithmetic
The headline price sounds small. Run it through realistic workloads:
| Monthly Search/Fetch calls | Monthly cost at $0.75/1,000 |
|---|---|
| 10,000 | $7.50 |
| 100,000 | $75 |
| 1,000,000 | $750 |
| 10,000,000 | $7,500 |
Now put it inside an agent. A research agent that performs one search every four steps, across a 40-step task, makes about ten retrieval calls per task. At 1,000 tasks per month, that is 10,000 calls — modest. At 100,000 tasks per month, it is one million calls and $750, before any model tokens are counted. Add a verification pass that fetches each source, and the number doubles.
The cost is not catastrophic. The problem is that it is unmodeled. Most teams priced their agents on tokens per task and assumed retrieval was free because it had been free since launch. A budget built on that assumption is wrong by a constant factor that scales with agent volume, not with answer quality.
Who gets hit hardest
Four workloads will feel the change immediately:
- Long-horizon research agents. They search to ground claims and fetch to verify them. Retrieval is not an optional enhancement; it is the workflow.
- Monitoring and watching agents. Scheduled agents that check sources hourly or daily generate calls whether or not anything changed. The cost is flat and relentless.
- Evaluation harnesses. Running a benchmark across thousands of tasks with web grounding multiplies retrieval calls by the size of the evaluation set — and evaluation is exactly when you least want to optimize away the tool.
- Multi-agent systems. If a planner and three workers each search independently, retrieval costs scale with the agent count, not the task count.
The workloads that escape are those with cached context: agents that fetch a source once, store it, and reason over the stored copy. Caching turns a per-call cost into a per-document cost, which is the only structural defense.
A checklist before the deadline
The deadline is four days away from the date of this article. Five actions are worth doing now.
Inventory the call sites. Find every place a server-side search or fetch tool is enabled: Harness configurations, Claude Code and Codex setups, internal agent frameworks. The free period has encouraged broad enablement; assume more is switched on than you remember.
Measure current volume. Log calls per task, per agent, and per day for one week. You cannot budget a meter you have not read. If logging does not exist, add it before the meter starts.
Separate necessary from exploratory retrieval. A search that establishes a fact is necessary. A search that rephrases the same question is exploratory. The second category is where budgets leak, and it is usually the majority.
Add caching at the tool boundary. Cache search results by normalized query and fetch results by URL, with a short TTL for news and a long one for stable documents. This is the highest-leverage change and it does not depend on the provider.
Set per-task and per-agent caps. A hard ceiling on retrieval calls per task converts an open-ended cost into a bounded one. Pair it with a graceful degradation path: if the cap is hit, return the best available answer with a confidence note rather than failing the task.
The strategic reading
DeepSeek is not the first lab to meter retrieval, but the move fits a pattern worth naming. Model inference is becoming commoditized faster than the surrounding services. Token prices have fallen repeatedly; the differentiators are shifting to orchestration, memory, retrieval, and evaluation — the layers where usage is harder to compare across providers and therefore easier to monetize.
For buyers, that means the unit of comparison is no longer “price per million tokens.” It is cost per completed task, including every metered dependency the task touched. An agent that appears cheap per token but calls search ten times per task can be more expensive than a model with higher token prices and a tighter retrieval loop.
For publishers, there is a second-order effect. Fetch calls are how agents read the web. When retrieval is free, agent traffic is invisible and unpriced; when it is metered, the economics of being read by agents becomes a question for both the agent operator and the site being fetched. That debate is already running — Wikipedia’s 2026 outage and the bot-traffic argument showed how quickly it can become contentious.
What to do on November 1
Do not wait for the invoice to teach you the shape of the new cost. Before the free window closes:
- confirm which tool calls are actually billed and which remain free;
- add retrieval-call logging if it is missing;
- cache aggressively at the tool boundary;
- cap calls per task;
- re-run your cost-per-accepted-task calculation with retrieval included.
The model release was the headline. The retrieval meter is the part that changes how you operate.
Bottom line
Search and Fetch were free long enough to become invisible. Starting October 31, they are $0.75 per 1,000 calls, and for agent workloads that is a second meter running alongside tokens. The teams that budget it deliberately will absorb the change; the teams that discover it on an invoice will spend the next quarter explaining a cost line nobody owns.