DeepSeek now sells agent web search for $0.75 per 1,000 calls, about a tenth of Tavily's basic search price, and the price only works for you if your agents' queries may leave your jurisdiction. The service runs on DeepSeek's hosted platform, whose privacy policy says it collects, processes and stores personal data in the People's Republic of China. You can download the new model that shipped alongside it under an MIT licence and run it anywhere. The search index stays in DeepSeek's hosted service.
Free calls end soon. Server-side search through Claude Code, Codex and DeepSeek's own harness is free until billing starts at 00:00 Beijing time on November 1, 2026 (16:00 UTC on October 31). Any team that wires it in during the free window should do the cost and data-path math first.
What DeepSeek Shipped on October 17
DeepSeek released two things today: a new checkpoint called DeepSeek-V4.1-Flash-1017 and a paid Search & Fetch API open to every API user. The release note describes the model as scoring higher, finishing tasks faster and writing shorter outputs than the previous Flash version. It says the search engine behind the API serves "hundreds of millions of requests daily worldwide," a figure DeepSeek has not had anyone verify.
The model card lists 763B total parameters, a 1M-token context window and an MIT licence covering both the code and the weights. On the hosted API, the pricing page lists the model as deepseek-flash at $0.30 per million input tokens on a cache miss and $1.20 per million output tokens at peak, half that off-peak.
One line on that pricing page matters more than the benchmarks. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but they are now served by the 1017 checkpoint. If your production code calls deepseek-v4-flash, the model behind it changed today without a version bump on your side. DeepSeek did the same thing with the 0731 checkpoint on July 31, and the advice is the same: re-run your evals before you trust last month's numbers.
How $0.75 Compares With Tavily, Exa, Brave and Perplexity
On list price, DeepSeek is the cheapest of the five search APIs compared here, and the gap is widest against Tavily. Both endpoints, /v1/tools/search and /v1/tools/fetch, cost $0.75 per 1,000 successful requests. DeepSeek bills only HTTP 200 responses that return results, and asking for up to 50 results does not change the price.
Here is the per-1,000-call search price for each vendor, checked on each vendor's live page on October 17, 2026:
| Vendor | Search, per 1,000 calls | Page content | Zero data retention |
|---|---|---|---|
| DeepSeek | $0.75 | Fetch $0.75 per 1,000 pages, or inline with need_content |
Not stated |
| Perplexity | $5.00 standard, $1.00 Fast Search | Not priced separately | Not on the pricing page |
| Brave | $5.00 | LLM Context endpoint, not priced separately | Enterprise plan |
| Exa | $4 instant, $7 auto | $1 per 1,000 pages | Enterprise plan |
| Tavily | $8 basic, $16 advanced | Extract: 1 credit per 5 URLs | Not on the pricing page |
Sources: Perplexity, Brave, Exa, and Tavily's pricing at $0.008 per pay-as-you-go credit, with basic search at one credit and advanced at two per its credit table.
Take a common research step: one search, then three page fetches. On DeepSeek that costs $0.003, or $0.00075 if you set need_content and skip the fetches. On Tavily it costs about $0.0128 (a basic search plus three-fifths of a credit for extraction). On Exa with auto search and text contents it costs $0.010. Run a million of those steps a month and the bill is roughly $3,000 on DeepSeek, $10,000 on Exa and $12,800 on Tavily.
The strongest case against switching is Perplexity's Fast Search at $1.00, which closes most of the gap with a US vendor. Throughput is the other catch. DeepSeek caps each endpoint at 5 queries per second per account, while Brave's Search plan lists 50. Five per second is 432,000 calls a day at the ceiling, enough for one team's coding agents and too few for a customer-facing fleet. Nobody has published an independent quality comparison of DeepSeek's index yet, so the price is the only number you can check. Our comparison of Tavily, Exa, Brave and Perplexity found page text drives the bill more than the search call, and DeepSeek prices both low.
Where Do Your Agent's Queries Go?
Every query and fetched URL goes to DeepSeek's hosted service, under terms governed by Chinese law. The Open Platform terms, effective April 29, 2026, say they are "governed by the laws of the People's Republic of China in the mainland," and they hand personal data handling to the privacy policy. That policy lists "User Input" among the data used to "train and improve our technology," and you opt out by emailing privacy@deepseek.com. The API terms do not say whether DeepSeek trains on API inputs, and the search guide says nothing about retention or region.
A search query sounds harmless until you read what a coding agent sends. It pastes stack traces, internal package names, customer identifiers in error messages, and the name of the vendor you are evaluating. A research agent sends the acquisition target it is looking at. The query log is a map of what your company is working on, even when the pages it fetches are public.
The steel-man for using it anyway is real. An agent that looks up public library docs for open-source work has little to leak, and DeepSeek's price makes high-volume grounding cheap. But you have to enforce that boundary in code, at the tool level, because the agent will not respect it on its own. The Fed, OCC and FDIC questions on bank AI agents are a preview of what a supervisor will ask about this data path. Our data residency buyer's guide covers the in-region options for the model side.
Does the DeepSWE Lead Hold Up?
DeepSeek's headline benchmark is real but narrow: the model leads DeepSWE v1.1 in DeepSeek's own table, against last-generation closed models. The model card shows 77.0 on DeepSWE v1.1, ahead of GPT-5.6 Sol at 73.0 and Opus 5 at 74.0, measured with the mini-swe-v2 harness. The table does not include GPT-6.1 Sol or Claude Opus 5.5, the models a buyer would actually compare it with.
Even against Opus 5, the lead is on one coding test. Opus 5 scores higher on Terminal-Bench 4.0 (51.8 to 42.3, run in the Claude Code harness), ProgramBench (37.0 to 29.1), NL2Repo-Bench (75.3 to 68.2) and HLE without tools (56.3 to 42.6). DeepSeek's checkpoint wins on the agent-workflow tests: Automation-Bench (60.5 to 50.3), Agents' Last Exam (35.4 to 28.6) and Apex Bench (72.1 to 65.8). Those are vendor-reported numbers on benchmarks DeepSeek picked.
At $1.20 per million output tokens at peak, the model does not need to beat the frontier to be useful. It needs to clear your own task suite at a cost per task you can live with. Our Together vs Fireworks vs Groq comparison showed how much tool-call accuracy varies across hosts serving the same open weights, so test on the host you will use.
Self-Hosting Gets You the Model Without the Index
The MIT weights let you keep inference inside your own boundary, but the cheap search only exists on DeepSeek's API. MIT carries no field-of-use limits, unlike the critical-infrastructure ban in Meta's Muse Spark Open licence. At 763B parameters with FP8 tensors, it is a multi-GPU deployment, and the card's guidance amounts to an SGLang container launched with --gpus all.
A self-hosted setup pairs the open model with a search vendor you can sign a data agreement with. Brave and Exa both list zero data retention, but only on their enterprise plans, so the clean data path costs more than $0.75 and requires a sales call.
If you use Claude Code against DeepSeek's Anthropic-compatible endpoint, note how it maps models: any claude-sonnet or claude-haiku name routes to deepseek-flash, any claude-opus name to deepseek-v4-pro, and cache_control is ignored. A developer who sets ANTHROPIC_BASE_URL on a laptop sends code, prompts and search traffic to the same place. Check for that variable in your managed developer environments.
What to Do Before Billing Starts October 31
This Week:
- Search your codebase and configs for
deepseek-v4-flashandapi.deepseek.com. Anything that matches is now running the 1017 checkpoint. Re-run its eval suite and compare against your last baseline. - Grep developer machine images and CI secrets for
ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic. Decide whether that routing is allowed, then write the decision into your developer AI policy. - Ask your privacy lead to rule on one question: may agent search queries go to a service governed by PRC law that stores data in China? Get the answer in writing before the free window makes it a default.
Before October 31:
- If the answer is yes for some workloads, add search routing at the tool layer. Public-docs lookups go to DeepSeek; anything touching customer data, deal work or internal code goes to a vendor under contract.
- Log a week of your agents' real search volume. Price it at $0.75, $5 and $8 per 1,000, and check whether 5 queries per second covers your peak.
This Quarter:
- If you want the model without the data path, stand up the MIT weights on your own GPUs or a US host. Benchmark them against Opus 5.5 or GPT-6.1 Sol on your own tasks, since DeepSeek's table did not.
- Get enterprise quotes with zero data retention from Brave and Exa so you know what the compliant search path costs.
The Bottom Line
DeepSeek has moved the list price of agent web search down by close to an order of magnitude against Tavily. The catch is the one that comes with its models: the open weights are portable and the hosted service is not, and the price increase DeepSeek flagged in August shows a launch price can change after you build on it. Teams that route by data sensitivity can take the cheap calls where the queries carry nothing worth protecting. Before billing starts October 31, write down which of your agents' queries are allowed to reach DeepSeek's servers.
Continue Reading
- Tavily vs Exa vs Brave: The Page Text Costs More Than the Search
- DeepSeek Swapped the Model. Your Eval Didn't Notice.
- DeepSeek Will Raise Prices. Your Ceiling Is Already 4x.
- LLM Data Residency: Which Providers Actually Keep Data In Region
- Together vs Fireworks vs Groq: Price the Tool Call, Not the Token
- Meta's Muse Spark Open License Bans Running Critical Infrastructure