Anthropic adds native web search to Claude API and Claude Code

Giving Claude real-time web access used to mean stitching together third-party search APIs, managing scrapers, and cleaning up messy HTML yourself.

Anthropic just made that whole pipeline obsolete. The company launched native web search across the Anthropic API, bringing live internet access and automatic citations directly into Claude models and its Claude Code tool.

Autonomous research with built-in guardrails

This is not just a basic search plugin. When developers enable the web search tool via the Messages API, Claude uses its own reasoning to decide if a prompt actually needs fresh data.

If it does, Claude formulates targeted queries, pulls relevant results, analyzes the content, and writes out a response backed by source citations.

Even better, Claude can run progressive searches. It can take findings from an initial query, refine its angle, and execute follow-up searches to conduct light research before answering. Developers can cap this agentic loop using the max_uses parameter to prevent runaway queries.

For enterprise teams worried about prompt leaks or sketchy sources, Anthropic added domain allow lists and block lists, alongside organization-level toggles.

The feature is available now for Claude 3.7 Sonnet, the upgraded Claude 3.5 Sonnet, and Claude 3.5 Haiku. It costs $10 per 1,000 searches, plus standard model token fees.

Early adopters are already on board. Quora is plugging the feature into Poe, calling it fast and cost-effective. Adaptive.ai noted that Claude's research agent behavior consistently beat out alternative search tools they tested.

Why it matters

For builders, this removes one of the biggest friction points in building production AI agents. Maintaining custom retrieval infrastructure for live financial updates, court records, or developer docs was a massive headache. Turning it into a simple API parameter saves weeks of plumbing.

Claude Code users get an immediate win too. Coding assistants notoriously struggle with zero-day library updates and breaking framework changes. Having Claude query the latest documentation on the fly solves hallucinated syntax errors before you even hit save.

We think Anthropic handled this cleanly by leaning into agentic reasoning rather than a rigid single-turn lookup. The combination of domain filtering and strict usage limits makes it viable for production apps with strict budgets. The only real question left is reliability: developers building deep research workflows will quickly test whether Anthropic's underlying search engine can consistently crack tricky, paywalled, or heavily dynamic web pages.