DeepSeek Search and Fetch: opening agent grounding APIs before the October 31 meter starts

Introduction

Flash-1017 improves the model. Search and Fetch change how agents reach the live web without bolting on a separate search vendor. For teams already routing Claude Code or Codex through DeepSeek, the free window is a two-week experiment budget that disappears on a hard UTC deadline.

Primary sources: DeepSeek change log (2026-10-17), news post, and Models and pricing (Search and Fetch table). Adjacent: Flash-1017 Pareto guide, MCP EMA, and Vercel MCP Events.

What Search and Fetch are

• Standalone APIs billed per successful request, separate from model token charges.
• Backed by DeepSeek’s in-house search stack that DeepSeek says already serves hundreds of millions of daily requests.
• Documented as multilingual, freshness-aware, and filtered against spam, GEO manipulation, and low-quality AIGC.
• Usable directly and via server-side tools in selected agent harnesses.

How this pairs with Flash-1017 and DeepSeek Harness

DeepSeek shipped Search/Fetch on the same day as Flash-1017 and asked harness users to move to DeepSeek Harness 0.2.1-rc.1 or later. The 0.2.1-rc line adds Search/Fetch settings alongside 1017 support. That is the intended stack for DeepSeek-native agent loops: newer Flash checkpoint, optional web tools, preview harness.

Claude Code and Codex users can stay on those products and still use DeepSeek as the model backend, then enable server-side Search/Fetch where the integration exposes it. Do not assume every third-party harness gets the free window. Confirm against DeepSeek’s pricing notes before promising zero search cost to stakeholders.

What this means for developers

• Add Search/Fetch call counts to agent telemetry next to token usage.
• Cache Fetch results for stable URLs inside a task to avoid repeat billing after November 1.
• Keep MCP browser or docs servers for authenticated internal systems; use DeepSeek search for public web grounding.
• Retest citation and hallucination gates when agents gain live search.
• Upgrade DSH to 0.2.1-rc+ before enabling web-search settings in production previews.

What this means for businesses

• Grounding spend becomes a second FinOps line item beside tokens.
• The free window is a calendar-driven pilot, not a permanent subsidy.
• Governance should decide which agent roles may call public search and what data may leave the VPC.
• Compare DeepSeek Search/Fetch against existing Bing, Google, Brave, or MCP search connectors on relevance and cost per completed task.

Two-week pilot plan

1. 01
   Pick one grounded workflow

   Docs Q&A, competitive research, or dependency advisory with clear acceptance tests.
2. 02
   Enable Search/Fetch in one harness

   DSH 0.2.1-rc+, Claude Code, or Codex on a DeepSeek API key with spend alerts.
3. 03
   Measure during the free window

   Log successful Search/Fetch counts, task success, and latency through October 31.
4. 04
   Set November budgets

   Convert observed call volume into $0.75/1k line items and approve or disable tools.
5. 05
   Decide dual-stack needs

   Keep MCP search for private corpora; keep DeepSeek search only where public web quality wins.

Conclusion

Search and Fetch turn DeepSeek from a model endpoint into a grounding platform for agents. The useful move is a measured pilot inside the free harness window, then an explicit November budget. Capability without metering becomes surprise spend the week after October 31.

Sources: api-docs.deepseek.com — updates ; api-docs.deepseek.com — news261017 ; api-docs.deepseek.com — pricing ; GitHub