"Can Claude look things up?" is one of the first questions enterprise teams ask, and on Vertex AI the answer needs precision. Two different concepts get conflated: Grounding with Google Search, a Vertex AI platform feature most prominently documented for Google's own Gemini models, and Anthropic's web search tool, a server-side tool built into the Claude API. For Claude on Vertex, the documented, supported path is the second one — Anthropic's web search tool, in its basic variant. If you're evaluating Grounding with Google Search specifically for Claude, confirm current support in Google's official documentation rather than assuming Gemini-oriented grounding features carry over; Google does maintain a dedicated web search page under its Claude partner-model documentation, which is the authoritative place to check.
What Claude's web search tool does on Vertex
Anthropic's web search is a server-side tool: you declare it in the tools array of a Messages request, and the platform executes searches on the model's behalf — no search infrastructure on your side, results flowing back into the conversation with citations. On Vertex AI, the availability matrix is specific: the basic variant, web_search_20250305, is supported; the newer web_search_20260209 variant with dynamic content filtering is not. This is a genuinely useful gap to know before you commit, because prompts and filter configurations tuned against the newer first-party variant won't port unchanged.
Pricing on Vertex matches the first-party rate: Google's pricing page lists web search at $10 per 1,000 searches, on top of standard token costs. Search usage is also metered in quota terms — Google's partner-model quota documentation includes web-search-specific quota metrics (for example, UsOnlinePredictionWebSearchRequestsPerProjectPerPublisher on the us multi-region endpoint), so a search-heavy application should size those quotas alongside its token quotas.
How this differs from the direct API
| Capability | First-party Claude API | Claude on Vertex AI |
|---|---|---|
Web search (basic, web_search_20250305) | Available | Available |
Web search with dynamic filtering (web_search_20260209) | Available | Not supported |
| Web fetch tool (retrieve a specific URL) | Available | Not supported |
| Search pricing | $10 per 1,000 searches | $10 per 1,000 searches |
The web fetch row deserves emphasis. On the first-party API, search and fetch pair naturally: search finds candidate pages, fetch pulls a full page into context. On Vertex, web fetch is not available — Google's documentation states the general rule plainly: Claude models on the platform support client tools, but server tools are not supported, with web search functioning as the notable exception carved out in the Claude partner-model docs. If your application needs to read specific URLs, you retrieve the content yourself and pass it inline — the same client-side pattern described in the Files API workaround article.
Configuration steps
Setup rides on the standard Vertex prerequisites — model enabled in Model Garden, roles/aiplatform.user for the caller, ADC configured. Then declare the tool in the request:
from anthropic import AnthropicVertex
client = AnthropicVertex(project_id="my-gcp-project", region="global")
message = client.messages.create(
model="claude-sonnet-5",
max_tokens=1024,
tools=[{"type": "web_search_20250305", "name": "web_search"}],
messages=[{"role": "user",
"content": "What changed in the EU AI Act guidance this quarter?"}],
)Responses interleave search-result content blocks with the model's text, including citation information — Vertex supports citations and search-results content blocks. Budget-wise, remember each search counts against the $10-per-1,000 meter regardless of how many results it returns.
Governance teams should also note that enabling web search changes your data-flow picture: the model now formulates queries derived from user prompts and sends them to an external search service. If your deployment runs inside a VPC Service Controls perimeter or under strict egress rules, review whether server-executed search fits your policy before turning the tool on, and confirm the current behavior with your provider's documentation.
Where to go next
The full availability picture is in the feature gaps overview. For the tool-use loop that client-side retrieval rides on, see tool use with Claude on Vertex AI.