Lab 031 · Editorial lab

Listed isn’t available. Call the model before you plan on it.

By Keith Townsend · November 19, 2026

Every cloud lists the frontier models, and each announces its own as generally available (GA). I tried to call the Clinic’s readers through each cloud’s own model API, on accounts I already use. The catalog and the GA label said nothing about whether my account could call the model.

What this covers: accounts I already use on each cloud, US regions, November 18 and 19, 2026. On Azure, a pay-as-you-go subscription I’ve had for years that had never used Foundry, Azure’s model service. On AWS, my lab account. On Google Cloud, two projects with production and test workloads. One call per reader on each route, a quota read against the Clinic’s rate, and one quota request where a cloud offered it. That shows whether a model is callable and what limit the cloud states, not sustained throughput. API access to hosted models only. It doesn’t cover Anthropic’s own API, which is how the Clinic has always reached Claude, or an Enterprise Agreement or committed-spend account. The open question: what does an enterprise agreement start on?

Don’t Don’t read a GA announcement or a catalog listing as access. Microsoft-Decision-1.1 went GA on November 17 and was listed in every region I read. On my subscription its quota was 0 of 0, and Azure refused the deployment both times I tried.

Do Before you plan on a model through your cloud, make one real call and read its quota against your workload’s rate. Here that was a handful of reads and calls per cloud, a few hundred tokens in all, and it settled what the catalogs couldn’t.

Don’t Don’t assume your cloud will let you call a frontier model it resells. Claude Haiku 5.5 answered on Bedrock on the first call. Vertex AI returned a zero-quota error on both projects I tried and denied the increase request in seconds. Azure stopped the deployment for company details, with zero quota behind that.

Do If the reader has to move between clouds, test every cloud at your rate, not just for a reply. gpt-oss-120b was the only reader all three accepted, and only AWS and Azure stated a limit above the bar. Vertex AI’s limit for it didn’t read.

Self-funded editorial. Google Cloud is a client of The CTO Advisor LLC. Microsoft and AWS are vendors under test. Anthropic makes Claude, the control reader. An Anthropic model ran the scripted checks under my direction and drafted this record; the evidence is each cloud’s own API responses, kept raw. None of Microsoft, AWS, Google or Anthropic reviewed this page.

Microsoft announced Microsoft-Decision-1.1 as generally available (GA) on November 17. The next day I tried to deploy it on an Azure subscription I've had for years, and Azure refused it at 0 of 0 quota.

That's one model on one cloud. But it raised a question every enterprise architect should ask before an architecture review. When a cloud lists a model in its catalog, or announces it as GA, can your account actually call it?

On my accounts, often it couldn't.

What did I test?

The Clinic is my scheduling bench. Code runs a simulated clinic's booking conversations, and when it can't parse a patient's reply, it asks a reader model a closed question. The Clinic reaches Claude through Anthropic's own API, but plenty of enterprises can't do that. They buy through the cloud they already have an agreement with.

So I tried to reach the Clinic's readers through each cloud's own model service: Azure AI Foundry, Amazon Bedrock and Google Cloud's Vertex AI. The bar was the rate the Clinic actually runs, about 1,300 requests and 400,000 input tokens a minute. On each cloud I read the catalog, the entitlement and the quota, then made one real call per reader. Where a cloud offered a quota request, I filed one.

That's an access test, not a load test. One call tells you a model answers, and a quota read tells you what limit the cloud states, but neither tells you the cloud will sustain the bar.

Who got through?

Only one route cleared both checks: Claude Haiku 5.5 on Bedrock. It answered on the first call, with a stated limit of 10 million tokens a minute, which is AWS's default.

Google Cloud returned a zero-quota error for Claude on both projects I tried. Its quota API said the limit was eligible for an increase, so I filed one on the test project at the Clinic's rate, with no justification attached. It was denied within 12 seconds. That's an automated answer.

Azure stopped me earlier, because an Anthropic deployment requires your company's industry, name and country before the quota check runs. I didn't submit them, and the quota read behind that step shows zero for all 13 Anthropic models.

So Google gave me a form and an automated no. Azure gave me no form at all. They stopped me at different points, and neither got me Claude.

What about each cloud's own model?

This is where Azure stands alone. Amazon's Nova 2 Lite and Google's Gemini 3.8 Flash both answered on the first call. Microsoft-Decision-1.1 was refused at the smallest capacity twice, on November 18 and again on November 19.

The reason is Foundry's quota tiers. Foundry sorts subscriptions into tiers, Free and 1 through 6. Both of my subscriptions landed on Free Tier the day I first turned Foundry on. In eastus2, 56 of 92 GA models with a quota line have zero quota on Free Tier, and Free Tier can't request an increase.

The only reader all three clouds accepted was gpt-oss-120b. Even there, only AWS and Azure stated a limit above the bar, and Vertex AI's limit for it didn't read.

Wouldn't an Enterprise Agreement change this?

It might. I haven't tested one, and that's the open question this lab leaves.

Microsoft's documentation says Foundry's starting tier weighs the customer's relationship with Microsoft, and an agreement is exactly that. My subscriptions are pay-as-you-go and had never used Foundry. So the result shows years of other Azure spend didn't lift the starting tier. It doesn't show what an agreement starts on.

What doesn't depend on the agreement is the check, and it costs almost nothing. It took a handful of reads and calls per cloud, a few hundred tokens in all, and it settled what the catalogs couldn't. Why would you find this out after the architecture is approved?

So what?

A catalog listing tells you a model exists, not that your account can call it. Before you plan on a model through your cloud, make one real call and read its quota against your workload's rate. If the reader has to move between clouds, test every one at your rate, not just for a reply.

I came in ready to write a Microsoft story, and on Decision-1.1 it still is one. On Claude it isn't, because two of three clouds gated it on my accounts. Listed isn't available. Call the model before you plan on it.

*Disclosure: Google Cloud is a client of The CTO Advisor LLC. Microsoft and AWS are vendors under test. Anthropic makes Claude, the control reader, and an Anthropic model drafted this essay from the lab's record.*

The numbers

The bar
From the three Claude Haiku 5.5 arm E cells at 16 parallel sessions: 1,136 to 1,301 requests and 329,000 to 382,000 input tokens a minute, about 310 tokens a call, 25 to 29 million tokens a cell, 70 to 81 minutes each.
About 1,300 requests and 400,000 input tokens a minute
Claude Haiku 5.5 on Amazon Bedrock
Equal to AWS’s default, as are Sonnet 5.5 (6 million) and Opus 5.5 (30 million). Bedrock states no requests-per-minute quota for Haiku 5.5, only tokens. One call, first try.
Answered; stated limit 10 million tokens a minute
Claude Haiku 5.5 on Vertex AI
Quota exceeded for anthropic-claude-haiku at the global endpoint, on the test project and then on the production project. The quota read beforehand returned no value for any Anthropic model, which turned out to mean zero. The model isn’t offered in us-east5 (404).
Zero-quota error on both projects
The Vertex AI quota request
Filed through the Cloud Quotas API on the test project for 1,500 requests, 500,000 input tokens and 100,000 output tokens a minute, global, at 10:55 AM Central on November 19. All three limits read eligible for an increase before filing. No written justification was sent. Denials are stamped 4 to 12 seconds after filing.
Denied within 12 seconds
Claude Haiku 5.5 on Azure AI Foundry
Anthropic deployments require company details (industry, organization name, country) before the quota check. I didn’t submit them. The quota read shows zero for all 13 Anthropic models.
Refused at deployment
Azure Foundry’s starting tier
Foundry sorts subscriptions into quota tiers, Free and 1 through 6. Both were set to Free when the Foundry provider was first registered, November 18. In eastus2, 56 of 92 GA models with a quota line have zero quota, including all 13 Anthropic models and 41 of 50 OpenAI models. Free Tier can’t request an increase, and there’s no API to file one.
Free Tier on both subscriptions
Microsoft-Decision-1.1
GA November 17. Refused at the smallest capacity twice: a Data Zone deployment (served within one geography) on November 18, and a Global deployment on November 19. Available capacity read zero in all 55 region and deployment-type entries.
Refused at deployment, 0 of 0 quota
The cloud’s own model on AWS and Google Cloud
Nova 2 Lite on Bedrock: 2,000 requests and 8 million tokens a minute. Gemini 3.8 Flash on Vertex AI: 50 million input tokens a minute and 5 billion a day; its requests-per-minute bucket returned no value, so its request limit isn’t known.
Both answered
gpt-oss-120b
Bedrock: 100 million tokens and 10,000 requests a minute. Azure: 5 million tokens a minute of quota, though the call ran on a deployment at the smallest capacity, 1,000 tokens a minute, deleted afterward. Vertex AI: answered at two endpoints; its quota read returned no value, so its limit isn’t known.
Answered on all three
Other open-weight models on Azure
Llama, Phi, Mistral and DeepSeek, each about 20 times under the bar. A run under that limit only finishes if the runner is paced to it: a 75-minute cell would take about a day. At 16 parallel sessions most calls would be refused.
20,000 tokens a minute
Gemma 4 26B and 31B
Not on Bedrock in three US regions. Public preview only on Vertex AI. On Azure, listed only for Microsoft-run compute on dedicated GPUs, not per token.
Not served GA per token anywhere

Where each layer belongs

Read for decision authority. Who makes the runtime decision at this layer, and can you see and override it? How to read this table

LayerAzure AI FoundryAmazon BedrockGoogle Cloud Vertex AI
Layer 2C · Reasoning
Agentic Infrastructure — The Reasoning Plane
Who decides whether the reader runs at my rate, not who decides inside it. On every route the cloud does, through its quota and entitlement policy, and the same model behind the same API got different answers. Bedrock granted Claude by default; Vertex AI set it to zero and denied the increase; Azure set Microsoft’s own model to zero with no way to request more. A granted limit is still the cloud’s decision. It’s Ceded where the call worked too, which is why the check belongs in the plan.
CededCededCeded

Sources

Where this fits

Last in this thread
What does a managed platform decide for you, and what do you run when you take it back?

Method and disclosure

Self-funded editorial. Reads came from each cloud’s own APIs and command-line tools: the model catalog, per-account restrictions and quota limits, and on Google Cloud the Cloud Quotas API for the increase request. Each reader got one call per route, “Reply with the single word: ready,” with 16 output tokens at most, plus a retry where the first failure was mine: Claude on Vertex AI was first sent to a region that doesn’t offer it, and the first Azure gpt-oss-120b call lacked a permission on my side. The reasoning models (gpt-oss-120b, Gemini 3.8 Flash) spent their 16 tokens thinking and returned no answer text; I count an accepted, billed request as callable. On Azure, deploying a model comes before a call, so each reader was deployed at the smallest capacity, called if it deployed, and deleted. Google Cloud quotas were read on the production project; the calls ran on the test project and Claude again on production; the quota request ran on the test project. Reads and the first Decision-1.1 deployment attempt ran November 18, the calls and the quota request November 19, 2026, all in US regions. The raw responses go in the lab detail download.

The essay on this page was drafted by Claude Opus 5.5, an Anthropic model, from the lab’s frozen structured record under this practice’s voice specification, checked by a deterministic gate against every figure in the record, and validated by the author before publication. The record is the canonical surface: where prose and record disagree, the record rules.

The structured recordThe lab’s canonical surface, as machines read it
Download the raw lab detail (Markdown)