DeepSeek Claude Code Web Search: official setup before you optimize prompts

Claude Code should point at DeepSeek's Anthropic-compatible base URL, not the OpenAI-format endpoint. This is the vendor-documented path for Claude Code and the only route that the official guide explicitly supports.

DeepSeek's guide keeps the main coding model on `deepseek-v4-pro[1m]`, while routing subagents to `deepseek-v4-flash`. That split matters because Web Search often creates extra summarization and follow-up requests that are cheaper on Flash than on Pro.

DeepSeek's guide says the API natively supports the Web Search feature in Claude Code. The wording matters: the search tool is triggered by the model when it decides a question requires web retrieval, rather than by a separate DeepSeek-only CLI.

The same paragraph also explains the billing implication. When Web Search runs, DeepSeek generates additional model requests to summarize retrieved search content. So the total cost is not only the visible final answer; it includes the extra retrieval-summary turns behind the scenes.

That makes Web Search an official feature, but not a free one. Treat it as a deliberate capability for current facts, docs lookup, and release verification, not as the default for every coding prompt.

Use Web Search for time-sensitive questions: latest package versions, current release notes, price changes, live incidents, or third-party docs that are not already in your repository. Those are the cases where stale memory is a bigger risk than extra tokens.

Do not lean on Web Search for tasks that are already local and deterministic. Refactoring a repository, explaining your own code, or editing files from a checked-out project should stay local unless the model truly needs a missing external fact.

A practical team policy is to let search stay available, but ask developers to keep prompts narrow: name the package, repo, version, or vendor doc you want rather than asking for broad market summaries.

DeepSeek's pricing page keeps V4 Flash cheaper than V4 Pro across cache-hit input, cache-miss input, and output tokens. That is why the official Claude Code page keeps Flash as the subagent lane.

Web Search increases cost because the model has to inspect, condense, and reason over retrieved content. The cleanest control is not 'never use search'; it is to keep the expensive main route reserved for work that truly needs Pro and let repetitive background tasks stay on Flash.

If a team is spending too much, first audit which prompts are triggering search, then tighten prompt scope, reuse stable system instructions for cache hits, and move routine summarization toward Flash-first substeps.

The biggest operational mistake is trusting the client label instead of the provider-side route. If the model name is wrong or unsupported, you can think you are on Pro while DeepSeek actually serves Flash-family behavior.

DeepSeek's API quick-start still treats `deepseek-chat` and `deepseek-reasoner` as compatibility aliases scheduled for retirement on July 24, 2026. That means older configs are not a stable way to preserve quality or search behavior.

After a real Claude Code session, check usage or provider logs where possible. If you are teaching a team, pin exact model strings in internal templates and remove old alias names completely.

If you are still cleaning up old model IDs, go straight to `/guides/deepseek-chat-to-v4-migration`. That page is the cleaner migration playbook.

If your next question is model choice rather than search behavior, use `/guides/deepseek-v4-pro-vs-flash` to decide when Pro truly earns its extra cost.

If you are comparing endpoint families, read `/guides/deepseek-openai-vs-anthropic-api-routing` before mixing Claude Code, SDK code, and agent tooling in one setup.