Vision — sending Claude an image alongside text and asking it to read, describe, or extract from it — is generally available for Claude models on Google Vertex AI, and it powers some of the highest-value enterprise use cases: invoice extraction, form processing, screenshot triage, and document review. The mechanics, though, have one platform-specific wrinkle that catches teams coming from Google's own models.
Inline base64 is the documented path
With Claude on Vertex AI, you embed the image directly in the request as a base64-encoded content block. Base64 is simply a way of writing binary bytes as text so they can travel inside a JSON payload. The AnthropicVertex client takes the same message shape as Anthropic's first-party API:
import base64
from anthropic import AnthropicVertex
client = AnthropicVertex(project_id="my-project", region="global")
img = base64.standard_b64encode(open("invoice.png", "rb").read()).decode()
msg = client.messages.create(
model="claude-sonnet-5", max_tokens=1024,
messages=[{"role": "user", "content": [
{"type": "image", "source": {"type": "base64",
"media_type": "image/png", "data": img}},
{"type": "text", "text": "Extract the vendor, date, and total."}]}])
print(msg.content[0].text)What about Cloud Storage URIs?
Teams used to Google's native models often expect to pass a gs:// URI and let the platform fetch the file. For Claude's online (real-time) requests, that shortcut is not available: Anthropic's platform documentation lists URL-based input sources and the Files API as unsupported on Vertex AI. In practice that means your application reads the object from Cloud Storage itself, base64-encodes the bytes, and sends them inline. The place Cloud Storage does appear as a first-class input is Google's own batch prediction feature for Claude, which takes JSONL input from Cloud Storage or a BigQuery table for asynchronous jobs — a different mechanism from the online API, covered in the batch workaround article.
Limits to design around
| Limit | Value on Vertex AI |
|---|---|
| Maximum size per image file | 5 MB |
| Maximum images per request | 100 |
| Maximum total request payload | 30 MB |
Note the interaction between the last two rows: base64 inflates file size by roughly a third, so a hundred moderately sized images will hit the 30 MB payload ceiling long before the per-image or per-count limits. If you are batching many photos into one request, resize and compress first. Common web formats such as PNG and JPEG are the safe choice; for the current definitive format list, check the official documentation. PDF input is also generally available on Vertex if your "image" is really a scanned document.
Cost implications
Images are billed as input tokens — the model "reads" a tokenized representation of each image, and larger images consume more tokens. Every image also counts toward the model's context window alongside your text, tool definitions, and output. Two practical consequences: first, downscaling images to the smallest size that preserves the details you need is a direct cost lever; second, a many-image request on a 200k-context model like Haiku 4.5 can crowd out room for the response. There is no per-image surcharge documented on Google's Claude pricing beyond token costs, but token counts vary with image dimensions, so measure a representative sample with the token counting API (supported on Vertex) before projecting costs.
Getting reliable answers out of images
Delivery mechanics aside, a few prompting habits determine whether vision workloads are production-grade or demo-grade. Put the image content block before the text instruction, as in the sample above, and be explicit about what to extract and in what format — pairing vision with structured outputs (generally available for Claude on Vertex) turns "read this invoice" into a schema-validated record your pipeline can trust. When a request carries several images, refer to them by position ("in the second image…") so the model and your reviewers are talking about the same thing. And keep a human-review lane for low-confidence cases: scanned documents arrive skewed, stamped, and coffee-stained, and asking Claude to flag unreadable fields explicitly is cheaper than silently ingesting wrong totals.
Where to go next
For end-to-end pipelines built on vision, see Document Processing: Contracts, Invoices, and Forms. If your payloads are bumping the 30 MB ceiling, Handling Large Payloads and Long Context Windows on Vertex AI is the companion piece.