# Lab 031 lab detail: listed isn't available

The substrate behind Lab 031: the bar, the ladder, every read and call on each cloud, the quota
request, what went wrong, and what the peer review changed. It's the raw detail, not the ruling;
the lab page carries the verdict. Nothing here is proprietary: the calls were one-word prompts,
and the rate bar comes from the Clinic's published cells (github.com/kltownsend/clinicbench).
Account, subscription and project identifiers, contact details and credential handling are left
out. Error messages are quoted as the clouds returned them.

## The question and the scope

Could the Clinic's reader be called through each cloud's own model API, on accounts I already
use, at the rate the Clinic runs? API access to hosted models only. Running a model on rented
GPUs was ruled out of scope; the GPU and compute quotas below are context, not result.

The accounts, all existing:

- **Azure:** two pay-as-you-go subscriptions, offer `PayAsYouGo_2014-09-01`, no spending limit,
  years of Azure use. Neither had registered the Foundry provider (`Microsoft.CognitiveServices`)
  before November 18, 2026.
- **AWS:** my lab account, us-east-1.
- **Google Cloud:** two billed projects in my organization, one production and one test, both
  with Vertex AI already enabled.

## The bar

From the three Claude Haiku 5.5 arm E cells at 16 parallel sessions, 48,000 conversations each:

| | per cell |
|---|---|
| reader requests per minute | 1,136 to 1,301 |
| reader input tokens per minute | 329,000 to 382,000 |
| tokens per call | about 310 |
| tokens per cell | about 25 to 29 million |
| wall clock | 70 to 81 minutes |

A limit under the bar doesn't stop a run if the runner is paced to it; it stretches it. At 16
parallel sessions most calls against a low limit would be refused. A daily token cap under about
30 million stops a cell outright.

## The ladder

1. **Catalog:** the model is listed and GA.
2. **Entitlement:** terms, forms and restrictions on this account.
3. **Quota:** the stated limit against the bar, and whether an increase can be requested.
4. **Call:** one real request. The first three passing doesn't mean the fourth will.

Every call used the same prompt, "Reply with the single word: ready," with 16 output tokens at
most. The reasoning models (gpt-oss-120b, Gemini 3.8 Flash) spent their 16 tokens thinking and
returned no answer text; an accepted, billed request counts as callable.

## Azure AI Foundry

**Tiers.** Foundry sorts subscriptions into quota tiers, Free and 1 through 6. Both subscriptions
were set to Free Tier when the Foundry provider was first registered: one on November 18 at
1:07 PM Central, the other at 6:21 PM. Tier upgrade policy on both: `OnceUpgradeIsAvailable`,
with no upgrade pending.

**Quota on Free Tier, eastus2.** 92 GA models have a standard quota line. 56 have zero on both
Global Standard and Data Zone Standard: all 13 Anthropic models, 41 of 50 OpenAI models (including both
text-embedding-3-large and ada-002), xAI's grok-4.7, and
Microsoft-Decision-1.1. The 36 with quota are mostly open-weight or small models (gpt-oss-120b,
Llama, DeepSeek, Mistral, Phi, Kimi, FLUX) plus gpt-5-mini, gpt-5.4-mini and grok-4-1-fast. One
quota unit is 1,000 tokens a minute.

| reader | stated quota | result |
|---|---|---|
| Claude Haiku 5.5 | 0 | refused at deployment: company details required |
| Microsoft-Decision-1.1 | 0 of 0 | refused at deployment, twice |
| gpt-oss-120b | 5,000,000 tokens/min | deployed at capacity 1, called, deleted |
| Llama, Phi, Mistral, DeepSeek | 20,000 tokens/min each | not called |

**Microsoft-Decision-1.1.** GA November 17. Microsoft's own models in eastus, eastus2,
swedencentral and westus3: Decision-1.1, MAI-Image-2.6 and MAI-Image-2.6-Flash have zero
everywhere; MAI-Thinking-1 (500, preview) and MAI-DS-R1 (20) have quota. Decision-1.1's available
capacity read zero in all 55 region and deployment-type entries. Deployment attempts, both at
capacity 1:

- November 18, Data Zone Standard (served within one geography): `InsufficientQuota`.
- November 19, Global Standard: `InsufficientQuota`. "The deployment requires 1 quota units, but
  only 0 quota units are available in Requests Per Minute - Microsoft-Decision-1.1. Current
  usage: 0 / 0 quota units."

Microsoft's documentation (Foundry Models quotas and limits, updated 2026-10-06, read November 18,
paraphrased): initial tier considers usage and the customer's relationship with Microsoft;
automatic upgrades mainly weigh Foundry consumption over time; some models need a higher tier
before they can be deployed. Decision-1.1's table starts at Tier 1 (2.16 million tokens and 2,160
requests a minute) with no Free Tier row. Free Tier subscriptions aren't eligible for quota
increase requests. No API exists to request quota; the tier API only sets the upgrade policy. The
Decision-1.1 GA announcement, the Foundry blog, the Learn how-to and the catalog page say nothing
about quota, tiers or eligibility.

**Claude Haiku 5.5.** Deployment at capacity 1, Global Standard: `InvalidModelProviderData`.
"ModelProviderData is required for Anthropic model deployments but was not provided. Please
provide all required fields: industry, organizationName, and countryCode." Not submitted.

**gpt-oss-120b.** Deployed at capacity 1 (1,000 tokens a minute), called once, HTTP 200, then
deleted; the account listed no deployments afterward. The first call failed with a permission
error on my side and was retried.

## Amazon Bedrock (us-east-1)

Entitlement read `AUTHORIZED` and `AVAILABLE` for every model checked.

| reader | stated limit | AWS default | result |
|---|---|---|---|
| Claude Haiku 5.5 | 10,000,000 tokens/min | same | answered (18 in, 4 out) |
| Claude Sonnet 5.5 | 6,000,000 tokens/min | same | not called |
| Claude Opus 5.5 | 30,000,000 tokens/min | same | not called |
| Nova 2 Lite | 2,000 requests, 8,000,000 tokens/min | | answered (53 in, 4 out) |
| gpt-oss-120b | 10,000 requests, 100,000,000 tokens/min | | answered (74 in, 16 out) |
| Gemma 3 27B | 100,000,000 tokens/min | | answered (16 in, 3 out) |

Bedrock states no requests-per-minute quota for Claude Haiku 5.5, only tokens, at both the
cross-region and global inference profiles. 23 of 1,123 applied quota values on this account
differ from AWS's defaults, some lower. The cross-model daily token cap is 310 million against a
700 million default, about ten cells a day. Gemma 4 isn't offered in us-east-1, us-east-2 or
us-west-2.

## Google Cloud Vertex AI

**Catalog.** Claude Haiku, Sonnet and Opus 5.5 listed GA in Model Garden; Gemini 3.5 through 3.8
Flash GA; gpt-oss-120b (pay per token) GA. Gemma 4 26B A4B per token is public preview; the GA
`gemma4` entry is a self-deploy container that needs a GPU.

**Quota read** (production project, November 18). Gemini 3.8 Flash: 50 million input tokens a
minute and 5 billion a day, stated explicitly. Every Anthropic bucket returned no limit value, as
did the default buckets for every per-model metric, including the generate-content request limit
Gemini traffic in that project runs under. Positive and unlimited (-1) values are stated
explicitly elsewhere in the same response, so an empty value was ambiguous: zero, or not
enforced per project. Only a call could tell.

**Calls.**

| reader | project | endpoint | result |
|---|---|---|---|
| Claude Haiku 5.5 | test | global | 429, quota exceeded |
| Claude Haiku 5.5 | test | us-east5 | 404, not offered in that region |
| Claude Haiku 5.5 | production | global | 429, quota exceeded |
| Gemini 3.8 Flash | test | global | answered (7 in, 13 thinking) |
| gpt-oss-120b | test | global and us-central1 | answered at both |

The 429 text, the same on both projects: "Quota exceeded for
aiplatform.googleapis.com/global_online_prediction_requests_per_base_model with base model:
anthropic-claude-haiku. Please submit a quota increase request." So an empty limit meant zero for
Claude, and didn't stop Gemini or gpt-oss-120b. Their request limits still aren't known.

**The quota request** (test project). Filed through the Cloud Quotas API at 10:55 AM Central,
November 19, base model `anthropic-claude-haiku`, global. All three limits read
`isEligible: true` before filing. No written justification; a contact email is required for any
increase and was supplied.

| limit | requested | granted | state | stamped |
|---|---|---|---|---|
| requests per minute | 1,500 | 0 | Quota request denied | 16:55:05 UTC |
| input tokens per minute | 500,000 | 0 | Quota request denied | 16:55:08 UTC |
| output tokens per minute | 100,000 | 0 | Quota request denied | 16:55:12 UTC |

Filed at 16:55:01 UTC, so each denial landed within 12 seconds.

## Context, out of scope: compute and GPU quotas

Read but not run; the scope ruling put self-hosting outside the test.

- **Runner (CPU).** Azure: 10 regional vCPUs and 3 spot in each of four US regions; the Container
  Apps and Container Instances providers weren't registered. AWS: 384 Fargate vCPUs, 1,920
  standard EC2 vCPUs. Google Cloud: 32 vCPUs across all regions, 1,000 Cloud Run jobs per region.
- **GPU.** Azure: every GPU family at zero in four US regions, several also restricted for the
  subscription. AWS: 768 on-demand G and VT vCPUs, 76 on-demand P vCPUs, spot zero. Google Cloud:
  GPUS_ALL_REGIONS zero, which caps every GPU in every region whatever the regional quota says.

## Gotchas

- **An empty quota value isn't a number.** Vertex AI's quota read returns no value for both zero
  and unenforced limits. Call the model.
- **Deploy before you call, on Azure.** A Foundry quota line isn't callable until a deployment
  holds it. A deployment at capacity 1 proves access, not the quota.
- **Registration starts the clock.** Foundry assigned the tier when the provider was first
  registered, not from the subscription's age.
- **The CLI's own quota.** The first Cloud Quotas filing was refused at a daily cap on increase
  requests charged to the command-line tool's shared project, not mine. Billing the call to my own
  project fixed it. It says nothing about Google's policy on Claude.
- **Region matters on Vertex AI.** Claude Haiku 5.5 is served from the global endpoint, not
  us-east5.

## Cost

A few hundred tokens across every call. The Azure test deployments existed for minutes and were
deleted. Nothing else was created on any cloud.

## What the peer review changed

Gemini (Antigravity, gemini-4-argon-high) and ChatGPT (gpt-5.6-sol) attacked the draft. Changes:

- "Served at the Clinic's rate" became "callable with a stated limit above the bar": one call
  shows access, a quota read shows the stated limit, neither shows sustained throughput.
- gpt-oss-120b is no longer offered as the portable reader at rate: Vertex AI's limit didn't read
  and the Azure call ran at capacity 1.
- The Google Cloud project split is disclosed, and Claude was called on the production project
  too. It returned the same 429.
- The Enterprise Agreement objection now concedes it's untested and drops the claim that account
  history didn't decide quota. Azure's subscriptions had no Foundry history; Bedrock's high
  defaults say nothing about clouds that start at zero.
- The placement table is read for decision, not portability, and the Layer 3 row, never run, is
  gone.
- Precision fixes: the 12-second window, Gemini's empty request bucket, the per-route call count,
  the date of the first Decision-1.1 attempt.
