TL;DR: From 12:00 Beijing time on Nov 18, 2026, DeepSeek no longer bills searches and page fetches that the model triggers itself. Those results now just count as tokens. Direct calls to the standalone Search and Fetch APIs stay at $0.75 per 1,000 successful requests. That changes which design is cheaper for a research agent.
What changed
| Route | Before Nov 18 | Now |
|---|---|---|
| Model-triggered web search / fetch | Per-call fee plus tokens | Tokens only |
| Your code calls the Search/Fetch API directly | $0.75 per 1,000 | $0.75 per 1,000 (unchanged) |
Source: AI Affairs report on the DeepSeek change. One tracker also lists a DeepSeek Flash output cut to $0.15 per 1M tokens (Overfeed price report); check DeepSeek's own pricing page before you budget on it.
The practical consequence
If your agent calls search through your own code, you pay the per-request fee. If you let the model call its built-in web tool, you pay only for the extra context tokens. For short snippets, the second route is now cheaper. For pages that return thousands of tokens, the token bill can exceed the old $0.00075 per call, so cap the fetched text.
Copy-paste: cap tool output so tokens stay small
def clip(text, max_chars=6000):
return text[:max_chars]
# after every web tool result:
messages.append({"role": "tool", "content": clip(result)})
log_cost(tokens_in=count(messages))
Worked example
Assume 10,000 research runs a day, 5 searches each. Old direct-call fee: 50,000 x $0.00075 = $37.50/day. Model-triggered: $0 in fees, but if each result adds 1,500 tokens at an assumed $0.035 per 1M input, that is 75M tokens = about $2.63/day. Your rates may differ; plug in your own.
Where to run it
A small always-on worker that clips and logs tool output is all you need. Get a Hostinger VPS here to host it cheaply.
Disclosure: this post contains an affiliate link. If you buy through it, TokenLean may earn a commission at no extra cost to you.
Top comments (0)