Anthropic made Claude Models on Google Cloud (Anthropic managed) generally available on October 22, 2026 — a second way to buy the same Claude model on the same Google Cloud invoice. Google documents both routes at the same price, with the same models on day zero. The difference is who is the data processor, which IAM role grants access, and which Claude Platform features exist. One day earlier, API version 2026-10-21 removed temperature, top_p and top_k from the Messages API, returning a 400 on every model.
same model, same price, same invoice
temperature, top_p, top_k · 400 on every model
inference_geo vs Google multi-region
The same Claude model, at the same listed price, on the same Google Cloud bill — and a different company operating the request. This table is the whole decision.
| Route | Who operates and processes | Endpoint and auth | New Claude features |
|---|---|---|---|
| Anthropic-managed | Anthropic operates the API and is the data processor under its own terms; Google Cloud gates access and bills | claude.googleapis.com/v1/projects/{project}/locations/global/workspaces/{workspace}/invoke Google OAuth bearer token; no x-api-key; roles/claude.viewer, claude.user, claude.admin per project | Day zero: Claude Managed Agents, skills, code execution, fast mode |
| Google-managed | Google Cloud operates the model as part of its Agent Platform and is the data processor under your existing agreement | aiplatform.googleapis.com :rawPredict, model named in the URL, anthropic_version in the body Google Cloud credentials with roles/aiplatform.user; VPC-SC, CMEK and residency available | New Claude Platform features after integration, not on launch day |
Sources: cloud.google.com · Expanding Claude availability on Google Cloud (same price, same billing, processor split) · platform.claude.com · Claude Models on Google Cloud (Anthropic managed) (endpoint, OAuth, IAM roles) · platform.claude.com · Claude Models on Google Cloud (Google managed) (Google-managed endpoint and version in body).
Google states the two options carry the same pricing. What moves is the processor, the compliance envelope, and how fast new Claude features arrive.
On the Anthropic-managed route, Anthropic operates the stack as the data processor on Google Cloud infrastructure under its own terms, while access runs through Google Cloud IAM, Cloud Audit Logging stays on, and spend consolidates under existing Google Cloud billing and eligible commitments. The Google-managed route keeps Google Cloud as processor under your existing agreement, with VPC Service Controls, CMEK and residency in the US, EU and APAC, and rolls new Claude features out as they are integrated. Billing is a single Google Cloud invoice on both — so a rate card cannot separate them, and neither can a per-token comparison.
The practical split: pick Anthropic-managed when you need the platform features at launch — Managed Agents, skills, code execution, fast mode on Opus 5.5 — and Google-managed when the processor must be Google or the workload needs VPC-SC, CMEK and regional residency.
API version 2026-10-21 applies to both Google Cloud routes and to every proxy in front of them. These are hard 400s, not deprecation warnings.
The sampling-parameter removal is the one that breaks silently at the edge: a proxy that forwards an OpenAI-shaped body still sends temperature, and Claude now rejects it even when the value is the default. The SDK majors that send 2026-10-21 by default — Python 2.0.0, TypeScript 1.0.0, and the rest — also carry their own breaking changes and move to Apache 2.0, so an SDK bump is a code change, not a patch.
The Google-managed route is not a subset by price — it is a subset by surface. These are absent there and present on the Anthropic-managed route.
Messages API, prompt caching, extended thinking, tool use, citations and structured outputs exist on both Google Cloud routes, so a classifier or a JSON-schema pipeline ports across without a rewrite. The divergence starts at files, long-running agents and the management APIs. On Amazon Bedrock, structured outputs for Claude became available on both the bedrock-runtime and bedrock-mantle endpoints on October 18, which closes the same gap on the third surface.
Once the price is equal, the premium comes from resolution and residency — and it is attached to the route, not the model.
| Route | Meter | What it changes |
|---|---|---|
| Claude API, Anthropic-managed | Standard Anthropic rates; inference_geo "us" adds 1.1x on input, output, cache writes and cache reads | Residency is opt-in per request and billed per token |
| Claude API, Anthropic-managed | Fast mode on Claude Opus 5.5: $8 / $40 per million tokens | Listed here and on Azure-hosted Foundry; not listed for the Google-managed route |
| Google-managed | Google Cloud rates; multi-region and regional endpoints add 10% over global | Regional endpoints cap at Claude Sonnet 4.6 and earlier |
| Amazon Bedrock | Europe and US inference profiles add 10% over global | Visible in the registry: claude-opus-5 at $5 / $25 global, $5.50 / $27.50 eu / us |
Sources: platform.claude.com · Pricing, caching and data residency (inference_geo 1.1x, 10% regional premium, fast-mode rate) · platform.claude.com · Claude Models on Google Cloud (Google managed) (regional endpoint limits).
The snapshot carries the models; it does not yet carry the split.
Snapshot cat_2026_09_22, observed September 22, 2026, records 373 Claude deployments across 17 providers. google-vertex-ai holds 39 of them, newest at Claude Opus 5, Claude Sonnet 5 and Claude Fable 5.1. Claude Opus 5.5 appears only on the anthropic provider, and the 5.5-era Sonnet and Haiku releases are absent because the snapshot predates them.
There is one google-vertex-ai provider id, so a Google-managed deployment and an Anthropic-managed deployment of the same model at the same price resolve to the same provider. Until the surface is modelled as its own provider, a registry query cannot answer which processor, which IAM role or which feature set applies — the report is the source of record for that split, not the catalog row.
Two changes in the October 19–26 window that move a route or a meter.
On October 20, the Responses API gained a 12-hour cache lifetime. A write at that lifetime costs 2.5x the input token price, against 1.25x for the default 30 minutes, while reads stay at 0.1x — and 0.05x on GPT-6.1 Sol. The 12-hour write only pays off when a stable prefix is reused hours apart; for a prefix read twice in a session, the 30-minute write is cheaper. Note the billing split: ttl_12h_tokens and ttl_30m_tokens are reported separately and an upgrade re-bills the prefix without credit for the earlier write.
On October 22 the Decisions API reached general availability with gpt-6-luna: up to 400 questions per request, input_json for structured evidence, images per question, and prompt caching on shared input with a 1,024-token minimum prefix. Pricing is unchanged at $0.10 per million input tokens with output tokens not charged — a meter a token-based routing model does not describe.
Price and invoice are identical, so the deployment record has to name the processor and the IAM boundary. A model id alone no longer identifies who handles the prompt.
Set anthropic-version explicitly on every call path, and remove sampling parameters before an SDK major upgrade sends 2026-10-21 by default.
1.1x for inference_geo on Anthropic-managed, 10% for multi-region or regional on Google-managed. Same model, different multiplier, different line item.
All official provider pages observed October 26, 2026 (Europe/Paris). ALLM catalog snapshot cat_2026_09_22 observed September 22, 2026. Prices are USD per million tokens as the provider states them.
Related live views: the Gemini 4 Argon price-cliff report, the October cheap-tier reset and the OpenAI vs Azure AI deployment comparison.
Deployment intelligence, explained
A provider record answers which company operates the API, who is the data processor, and how access is granted. Two routes can carry the same model at the same price on the same invoice and still differ on every one of those, so a price comparison returns no signal at all.
The request contract is the second axis. Claude on Google Cloud carries the Claude API body, which means an API version change applies to both Google Cloud routes and to every proxy in front of them. Pin the version, then read what the version removed — not the other way round.
We use technologies necessary for the website to operate. With your consent, we use PostHog and Google Analytics to measure pseudonymised audience data. Refusal does not affect access to the website.