DeepSeek Search and Fetch API: $0.75 Per 1,000 Calls Explained

DeepSeek now sells its own web search engine as an API, and it's priced to make every other provider look expensive. On October 17, 2026, DeepSeek opened up Search and Fetch endpoints to anyone with an API key: $0.75 per 1,000 successful calls for each one, billed only when a request actually returns something. The announcement rode along with the DeepSeek-V4.1-Flash-1017 model refresh, but the pricing is the part that's been passing around X and Hacker News all week.

I've spent a few days digging through the docs, the pricing pages of the competition, and the thread where people are already arguing about whether this is a steal or a liability. Here's what's actually going on, how to wire it into Claude Code or Codex, and when I'd reach for it instead of Tavily or Exa.

What is the DeepSeek Search and Fetch API?

It's two separate endpoints built on the same search engine that powers DeepSeek's consumer app, which DeepSeek says handles "hundreds of millions of requests daily worldwide."

• Search (POST https://api.deepseek.com/v1/tools/search): takes a query and returns up to 50 ranked results with titles, URLs, snippets, longer summaries, a 0-3 relevance score, and a 0-4 authority tier.
• Fetch (POST https://api.deepseek.com/v1/tools/fetch): takes one URL and returns the page as clean Markdown, ready to drop into a model's context.

Search supports the operators you'd expect from a real search engine: site: and -site: for domain filtering, exact-phrase matching with quotes, lang:zh,en for language, and after:/before: for date ranges. You can also pass need_content to pull full page text instead of just a snippet, and localization to bias results by IP.

One detail worth calling out: a few outlets reported a 5-requests-per-second limit, but DeepSeek's own docs say otherwise. The real account-level caps are 100 QPS for Search and 25 QPS for Fetch, with an additional per-domain limit on Fetch to keep any single site from getting hammered. That's generous enough for most agent workloads: you'd have to be running a pretty aggressive crawler to hit it.

How much does it actually cost?

$0.75 per 1,000 successful requests, for both Search and Fetch. "Successful" means HTTP 200 with a non-empty result. A failed request or an empty search doesn't get billed. That's a meaningfully honest pricing model; most providers charge for the attempt, not the outcome.

Here's where that lands next to everyone else selling the same thing:

So DeepSeek undercuts Tavily by roughly 10x and Anthropic's own built-in web search tool by over 13x. At a million agent search calls a month, you're looking at about $750 on DeepSeek versus $10,000 on Tavily's basic tier. That's not a rounding error, that's the difference between a search budget and a model budget.

One thing I can't verify yet: third-party accuracy benchmarks. An independent test of 300 queries put Exa Fast at 99.3% accuracy and Tavily Advanced at 93%, but DeepSeek's engine wasn't part of that run: it's too new. Price is confirmed. Quality is still "trust but verify."

How do I use it with Claude Code?

DeepSeek's API natively plugs into Claude Code's built-in web search tool, no separate client needed. Point Claude Code at DeepSeek by setting these environment variables (or dropping them into .claude/settings.local.json):

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API key>
export ANTHROPIC_MODEL=deepseek-flash

Once that's set, Claude Code decides on its own when a prompt needs a web lookup and fires the request through DeepSeek's Search tool automatically. You don't call the API directly. The model does.

How do I use it with Codex?

Codex needs deepseek-v4-pro specifically; web search isn't available on the flash tier inside Codex. Create ~/.codex/config.toml:

model = "deepseek-v4-pro"
model_provider = "deepseek"
preferred_auth_method = "apikey"

[model_providers.deepseek]
base_url = "https://api.deepseek.com/"
wire_api = "responses"
experimental_bearer_token = "<your API key>"

[features]
web_search_request = true

Flip web_search from "disabled" to "live", restart Codex, and it'll run searches server-side and feed the results straight back into the response.

Is it actually free right now?

For a couple more weeks, yes, but only through specific channels. Search and Fetch calls made via server-side tools in Claude Code, Codex, or DeepSeek's own Harness CLI are free until October 31, 2026, at 16:00 UTC. After that, standard per-call billing kicks in and draws from your regular DeepSeek balance, same pool as your token spend. If you're calling the Search or Fetch endpoints directly rather than through one of those three integrations, you're paying from day one.

What's the catch?

There's always one, and here it's data residency. Queries hit DeepSeek's infrastructure, which processes and stores data in China under Chinese law. DeepSeek uses inputs to improve its models by default: you have to opt out, not opt in. Italy and Germany have already taken regulatory action against DeepSeek over exactly this kind of data transfer.

For a side project or a personal coding agent, I don't think that's a dealbreaker. For anything touching customer data, health records, or an enterprise codebase with an NDA attached, it's a real non-starter regardless of how good the price looks. Your agent's search queries often leak more than you'd expect (stack traces, internal hostnames, customer IDs), and that context goes wherever the search provider's servers are.

Also worth knowing: the DeepSeek-V4.1-Flash-1017 model weights are MIT-licensed and you can run them yourself, but the search engine isn't portable at all. It only runs on DeepSeek's hosted platform. There's no self-hosted version, and there isn't going to be one. The value here is their index, not their code.

DeepSeek vs Tavily vs Exa: when does each one win?

Honestly, it depends on what you're optimizing for:

• Pick DeepSeek if you're cost-sensitive, building a side project or internal tool, and don't have data residency requirements. At $0.75 per 1,000 calls it's cheap enough to let an agent search liberally without watching the meter.
• Pick Tavily if you need an agent-tuned search layer with structured answer extraction and you're already paying for quality over cost: it's built specifically for RAG and agent pipelines, not general search.
• Pick Exa if raw retrieval accuracy matters more than price and you're doing semantic/neural search rather than keyword lookup. The third-party accuracy numbers back this up.
• Pick Brave or SerpAPI if data residency or "not China, not a frontier lab's training pipeline" is the actual requirement, and you're willing to pay $1.50-$5.00 per 1,000 for that.

I'll be testing DeepSeek's Search API against the agent workflows I run through DeepSeek V4.1 Flash 1017 over the next few weeks before the free window closes on October 31. If the SEO-spam filtering holds up the way early testers are claiming, this becomes the default search tool for anyone running cost-conscious coding agents, not because it's the best, but because at this price the gap barely matters for most use cases. If you're also comparing agent tooling costs more broadly, our roundup of the top AI tools for 2026 and the Claude Opus 5 pricing breakdown are good companion reads.

Watch