llmrelay vs Google Vertex AI

Claude Opus 4.7, per million tokens

Read from cloud.google.com/vertex-ai/generative-ai/pricing on 26 July 2026. Against the US or EU endpoint, llmrelay is 55% cheaper on input, not 50%.

The residency tax

This is the part worth understanding before you commit. If compliance requires that inference stay inside the US or the EU, Vertex's global endpoint doesn't qualify and you move to a multi-region endpoint — which is where the 10% premium lands. The residency requirement and the price increase are the same decision.

One gap worth flagging honestly: single-region Vertex pricing for Opus 4.7 isn't published. The per-region tables list Opus 4.6, Opus 4.5, Sonnet and Haiku. Only global, US multi-region and EU multi-region carry published 4.7 rates. If you need a specific single region, you'll have to ask Google.

Where Vertex is the right choice

• — You have a contractual data-residency obligation and need a documented US or EU endpoint.
• — You're already on GCP and want inference on the same invoice, under the same IAM and VPC-SC controls.
• — You're mixing Claude and Gemini in one application and want both in one console.
• — You have committed-use GCP spend to burn.

Where llmrelay wins

Most teams on Vertex aren't there for residency, they're there because the company already uses GCP. If that's you, you're paying $5.00 or $5.50 per million input tokens for a billing convenience. The same model IDs cost $2.50 here.

Setup difference is meaningful too: Vertex needs a GCP project, Model Garden enablement, quota, and ADC or service-account credentials. llmrelay needs a base URL and a bearer token.

Migration

# Before (Vertex — GCP auth, regional host, publisher model path)
gcloud auth application-default login
POST https://us-east5-aiplatform.googleapis.com/v1/projects/$PROJECT/locations/us-east5/
     publishers/anthropic/models/claude-opus-4-7:streamRawPredict

# After (llmrelay)
OPENAI_BASE_URL=https://api.llmrelay.dev/v1
OPENAI_API_KEY=sk-llmrelay-...
model: claude-opus-4-7

No project, no location string, no token refresh. Anthropic-native /v1/messages is supported if your code already speaks that format.

A note on token counts

Anthropic states that Claude 4.7 and later use a newer tokenizer producing roughly 30% more tokens for the same text. Same rate card, more tokens per document. Any cost model carried over from Opus 4.6 will understate spend on Vertex and here alike by about a third.