DeepSeek Prices Agent Web Search at $0.75 per 1,000 Calls

On October 17, 2026, DeepSeek paired two launches aimed squarely at agent builders: public Search and Fetch APIs backed by its in-house search engine, and an updated Flash model—DeepSeek-V4.1-Flash-1017—with MIT-licensed weights on Hugging Face and ModelScope. The search tools give any developer with a DeepSeek API key a priced path to live web results and Markdown page reads; DeepSeek reports better coding and agent benchmark scores for the refresh, which keeps the same deepseek-flash API name.

Key facts

What DeepSeek launched

DeepSeek's Oct. 17 post is a double release for agent builders. The Search & Fetch APIs are "now open to everyone," powered by an in-house engine that DeepSeek says serves hundreds of millions of requests daily worldwide. DeepSeek-V4.1-Flash-1017 simultaneously replaces the previous V4.1-Flash checkpoint on the website, app, and API.

DeepSeek calls the Flash update a Pareto improvement in performance and cost—better benchmarks, faster tasks, shorter outputs. It targets DeepSeek Harness (v0.2.1-rc.1+ recommended), Tencent CodeBuddy, and WorkBuddy, and uses reinforcement learning with self-feedback; demos include a Forbidden City 3D model, a web racing game, and a Godot Portal 2 recreation. The base V4.1-Flash line was described as a 552B-parameter MoE with 8B active input / 16B active output in DeepSeek's September launch post; DeepSeek's 1017 materials do not state a separate parameter count.

How Search and Fetch work—and what they cost

Per DeepSeek's Search & Fetch guide, Search (POST /v1/tools/search) accepts a 1–128 character query (with operators like site: and lang:) and returns ranked results with snippets, summaries, and optional page text. Fetch (POST /v1/tools/fetch) retrieves one URL and returns Markdown content plus a title.

Each endpoint costs $0.75 per 1,000 successful requests. Billing applies only when Search returns HTTP 200 with a non-empty results array or Fetch returns HTTP 200 with content—failures, empty searches, and timeouts are free. Result count (1–50) does not change the per-call price. Account limits are 100 QPS for Search and 25 QPS for Fetch (Fetch also has a per-domain QPS limit), separate from model concurrency. For another agent-search stack, see our coverage of Cloudflare's web search API and AI Gateway.

Free Search and Fetch through Claude Code and Codex until Oct. 31

When DeepSeek Harness, Claude Code, or Codex call Search and Fetch as server-side tools, those requests are currently free for DeepSeek API users. Billing starts Oct. 31, 2026, 16:00 UTC (00:00 Beijing time Nov. 1 on the pricing page)—after that, agent usage is metered like direct API calls.

What changed in V4.1-Flash-1017

The Hugging Face model card calls 1017 a minor update to DeepSeek-V4.1-Flash and publishes side-by-side scores against the previous Flash checkpoint. The table below reproduces those figures. All are DeepSeek-reported and have not been independently verified.

DeepSeek ran Terminal-Bench 4.0 with the Claude Code harness and DeepSWE v1.1 with the Mini-SWE-V2 harness.

DeepSeek's table also reports scores it lists for Claude Opus 5 (74.0 DeepSWE, 56.3 HLE) and GPT-5.6 Sol (73.0 DeepSWE). API clients still use deepseek-flash; legacy names route to 1017 at Flash pricing.

Token pricing and peak hours

Chat on deepseek-flash still uses token pricing: per Models & Pricing, cache-hit input is $0.003–$0.006/1M, cache-miss input $0.15–$0.30/1M, output $0.60–$1.20/1M (off-peak vs peak). Peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays.

Open weights under MIT

Weights ship under the MIT License on Hugging Face and ModelScope—the model card recommends a 1M-token context. That parallels other recent open releases such as Meta Muse Spark and StepFun Step 5 Preview, though hosting a large MoE locally remains a niche choice compared with the hosted API plus Search/Fetch bundle.

What it means for agent builders

DeepSeek is bundling Flash, first-party search, and harness hooks—less glue code than pairing a model API with a separate search vendor and HTML-to-Markdown pipeline. Queries and URLs pass through DeepSeek's systems; enterprises with residency requirements should review DeepSeek's privacy policy and terms (outlets such as Startup Fortune note storage of personal data in China—verify before production). Handle empty search results and documented Fetch errors (403/404) in agent loops, and stay under QPS caps to avoid 429s.

What to watch next

Mark Oct. 31, 2026 (16:00 UTC) for the end of free agent Search/Fetch, watch real-world retrieval quality after billing starts, and track harness releases (v0.2.1-rc.1+) and partner integrations. If 1017's gains hold in production, Flash plus $0.75-per-thousand search could become a default economical stack for web-grounded agents.

Sources

• DeepSeek API: DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real — Oct. 17, 2026
• DeepSeek API: Search & Fetch guide
• DeepSeek API: Models & Pricing
• Hugging Face: DeepSeek-V4.1-Flash-1017 model card
• DeepSeek: Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. — Sept. 10, 2026
• Startup Fortune: DeepSeek Opens Its Search and Fetch APIs at $0.75 per 1,000 Calls — Oct. 18, 2026
• PPC Land: DeepSeek prices web search for AI agents at $0.75 per 1,000 calls — Oct. 17, 2026