The verdict up front: DeepSeek opened its own search engine to developers this weekend — two JSON APIs, Search and Fetch, at $0.75 per 1,000 successful calls. That's roughly 10x cheaper than Tavily and 13x cheaper than Anthropic's native web_search tool. If your coding agents run through Claude Code, Codex, or DeepSeek Harness, you're using both for free until midnight Beijing time on November 1.
Quick housekeeping: no affiliate links in this one — links below are to docs and coverage, nothing sponsored. I write these breakdowns because pricing shifts like this change what it's sensible to build.
One honesty note before we go on: I haven't benchmarked these APIs myself. Nobody outside DeepSeek has had the chance yet — they launched days ago. Everything below comes from DeepSeek's docs and this week's coverage (TechNode's reporting is the most detailed so far), and I'll flag vendor claims as vendor claims. The prices are real. The quality is still an open question. That gap is kind of the whole point of this article.
What actually launched
Two endpoints on the DeepSeek API, using the same Bearer auth and api.deepseek.com base URL as the model API — your existing key just works:
Search — send a query, get back 1 to 50 structured results from DeepSeek's own search engine. Each result carries a short snippet (~100–200 characters), a longer model-oriented summary (~600–800 characters, written to be read by an LLM rather than a human), and an authority rating. JSON in, JSON out.
Fetch — send a URL, get back the page's title and content as Markdown. If you've ever wired Jina Reader into an agent pipeline, this is that job description, except it comes from the same vendor as your model and lands on the same invoice.
Both are capped at a reported 5 queries per second per account, and here's the detail I like most: only successful requests are billed. Failed calls, empty results, and timeouts cost nothing. Most search APIs bill per attempt — DeepSeek says it bills per useful result.
The context matters. The APIs landed alongside DeepSeek-V4.1-Flash-1017 on October 17 (MIT-licensed weights, 763B-parameter MoE, 1M context — DeepSeek claims 77.0 on DeepSWE versus 74.0 for Claude Opus 5, though those numbers are vendor-run and independent evals haven't landed yet). Cheap frontier-class model plus native retrieval: DeepSeek is now a full-stack agent platform, not just a model vendor.
The price table, because that's the story
Prices per 1,000 requests, from current pricing pages and this month's API comparisons:
| API | Price / 1k | What you get | Free tier |
|---|---|---|---|
| DeepSeek Search | $0.75 | Own index, structured results + summaries + authority score | Free until Nov 1 via agent harnesses |
| DeepSeek Fetch | $0.75 | URL → clean Markdown | Same |
| Cloudflare AI Gateway (Ceramic) | $0.25 | Resold third-party index, raw results | — |
| Serper | $0.30–$1.00 | Google results, snippets only, no page fetch | 2,500 queries (one-time) |
| Brave Search API | ~$5 | Own index, LLM Context endpoint | $5 credit/month |
| Exa | $4–$7 ($12–15 deep) | Neural search, page contents included | $10 credit |
| Tavily | $8 basic / $16 advanced | AI-native, citations, extraction | 1,000 credits/month |
| Anthropic web_search tool | $10 | Model decides when to search (+ token costs on results) | None |
| Bing Grounding (Azure) | Azure-metered | Grounded answers inside the model | Check Azure pricing |
Read it honestly: DeepSeek is not the absolute cheapest line on this table. Cloudflare's pass-through and Serper's volume tier undercut it on raw price. But Serper gives you snippets, not content — and DeepSeek's Fetch at $0.75 is the page-content story, which is the expensive half of most agent pipelines. Against the tools agent builders actually use for retrieval (Tavily, Brave, native model search), DeepSeek is an order of magnitude cheaper. Ten thousand searches a day costs about $7.50 with DeepSeek versus $80 with Tavily basic. That's not a rounding error; that's an architecture decision.
Who should switch
- You run a research or data-collection agent in a loop. If your pipeline is search → fetch → extract, repeated thousands of times a day, this is the single cheapest line item you can attack. "Failures are free" billing makes flaky-query workloads dramatically cheaper than APIs that charge per attempt.
- You're already a DeepSeek shop. Same key, same invoice, no new vendor review. The integration cost is near zero.
- Your coding agents run on Claude Code, Codex, or DeepSeek Harness. You're getting both APIs free until November 1. That's a no-excuse trial window — use it or lose it.
- You're prototyping RAG. Cheap retrieval changes what you can afford to prototype.
Who should wait
- You need Google-specific results. SEO tooling, rank tracking, anything where "what does Google say" is the product — Serper and friends still own that.
- You need compliance-friendly vendors. DeepSeek is a Chinese company; your queries and fetched pages go through their servers. If your threat model or your customers care about that, price doesn't settle it.
- You need proven result quality. This is the big one. Nobody has independently benchmarked DeepSeek's index for recall or precision against Google, Brave, or Bing. The price is verified; the quality is a vendor claim. Do not migrate a production pipeline on price alone.
- You need more than 5 QPS. The reported per-account rate limit is fine for development and modest workloads, tight for anything bursty at scale.
What I'd actually do
If I were building a research agent this week, I'd prototype on DeepSeek Search + Fetch today, run my own eval queries through it alongside whatever I currently use, and compare quality on my queries — not on anyone's benchmark. The free window (or $0.75/1k after it) makes the experiment nearly free. Then decide with data, not with a pricing page.
And separately: Anthropic quietly made Haiku 5.5 the default Haiku this month — 1M context at a reported $0.10/$0.50 per million tokens. The two stories rhyme. The cost of running agents is collapsing from both ends at once: the model and the retrieval. If your agent architecture was designed when search cost $8–10 per thousand, it's worth re-examining every assumption in it.
FAQ
How much do the Search and Fetch APIs cost?
RMB 5 — listed as $0.75 — per 1,000 successful requests, each. Failed requests, empty results, and timeouts are not billed.
When does the free window end?
Midnight Beijing time on November 1, 2026 (that's 16:00 UTC on October 31). Until then, server-side search/fetch calls made through DeepSeek Harness, Claude Code, and Codex are unbilled.
Do I need a new API key?
No. Existing DeepSeek API credentials work — same Bearer auth, same api.deepseek.com base URL.
What does the Search API return?
1 to 50 structured results per query: short snippets, longer model-oriented summaries, and an authority rating per result.
What does the Fetch API return?
A webpage's title and content converted to Markdown — built for feeding into agents and RAG pipelines.
Is it better than Tavily or Serper?
Cheaper, verifiably — about 10x less than Tavily basic per thousand calls. Better is unproven: no independent quality benchmarks of DeepSeek's search index exist yet. Treat price as fact and quality as a hypothesis you're going to test.
Are you running an agent that hits a search API in a loop? What are you actually paying per thousand — and does $0.75 change what you'd build? Genuinely curious what real-world agent search bills look like out there.
