Terminology first, because Google Cloud and Anthropic use different words for related ideas. "Context caching" is the phrase you'll often hear in Vertex AI conversations; Anthropic's feature for Claude is called prompt caching. It lets you mark a stable prefix of your request — a long system prompt, tool definitions, a reference document — so that repeat requests reuse the cached prefix instead of reprocessing it, cutting cost and latency. For Claude on Vertex AI, the question enterprises actually need answered is: is Anthropic's prompt caching available there, and does it behave like the direct API?
Availability: supported, with one gap
Prompt caching with both the 5-minute and 1-hour cache durations is generally available for Claude on Vertex AI — Google maintains a Claude-specific prompt caching page in its partner-model documentation, and Anthropic's availability matrix lists it as GA on Vertex. It also works on the global endpoint, so you don't have to trade caching for availability-optimized routing.
The gap: automatic prompt caching is not supported on Vertex AI. On the first-party Claude API, a single top-level cache_control field lets the platform manage cache breakpoints for you, and Anthropic recommends it as the default. On Vertex you use the explicit style instead — cache_control markers placed on individual content blocks. Functionally you get the same discounts; you just decide where the cache boundary sits.
from anthropic import AnthropicVertex
client = AnthropicVertex(project_id="my-project", region="global")
msg = client.messages.create(
model="claude-opus-4-8",
max_tokens=1024,
system=[{
"type": "text",
"text": LONG_POLICY_DOCUMENT, # stable prefix, reused often
"cache_control": {"type": "ephemeral"},
}],
messages=[{"role": "user", "content": "Does clause 7 apply here?"}],
)
print(msg.usage) # includes cache write/read token countsToken accounting and pricing
Google's Vertex pricing page lists the same caching multipliers as the direct API: writing a 5-minute cache costs 1.25x the base input rate, writing a 1-hour cache costs 2x, and a cache hit costs 0.1x. For Claude Opus 4.8 ($5.00 per million input tokens on the global endpoint) that means $6.25 per million tokens for a 5-minute cache write, $10.00 for a 1-hour write, and $0.50 for reads. The arithmetic follows: a 5-minute cache pays for itself after a single reuse, a 1-hour cache after two. Watch the usage object in responses — input tokens, cache-creation tokens, and cache-read tokens are reported separately, and all three still count toward the model's context window. Caching changes what you pay, not how much fits.
token_count metrics in Metrics Explorer.Configuration differences from the direct API, summarized
| Aspect | Claude API (1P) | Vertex AI |
|---|---|---|
Explicit cache_control breakpoints | Yes | Yes (GA, 5m and 1h) |
| Automatic prompt caching | Yes (recommended default) | Not supported |
| Cache read / write multipliers | 0.1x / 1.25x / 2x | Same |
| Cache diagnostics (beta) | Yes | Not available |
One more planning note: minimum cacheable-prefix sizes vary by model (Anthropic documents 1,024 tokens on Opus 4.8, and a lower 512-token minimum for Claude Fable 5 on the first-party API). Platform-specific minimums can differ — Anthropic notes Fable 5's minimum remains 1,024 on Bedrock, for instance — so verify the current figure for your model on Vertex in the official docs before assuming a short prefix will cache.
Where to go next
See the Vertex prompt caching how-to for breakpoint placement patterns, prompt caching costs for the economics across platforms, and platform feature gaps for the full availability matrix.