Research

What a federal agency can actually run: the AI authorization tables, read closely

Vendor announcements say a service is FedRAMP authorized. The published tables say something more specific, and in places the opposite of what the market assumes.

·

11 min

var(--variable-rpcX8dOAk)

Key findings

The authorization belongs to the cloud and the region, not to the service name. The same AI service carries materially different scope depending on where it is deployed.

Government is not always behind commercial. Four model families on Amazon Bedrock are authorized in AWS GovCloud and at DoD IL4/IL5 while carrying no authorization in AWS commercial East and West regions.

Guardrails can be less authorized than the model they guard. In Azure Government, Azure OpenAI is in audit scope through IL6 while Azure AI Content Safety and the Microsoft Foundry portal stop at FedRAMP High and IL2.

The baselines have been renamed. AWS now publishes FedRAMP Class C, formerly Moderate, and Class D, formerly High. Microsoft's tables still use FedRAMP High. Two vendors, two vocabularies, one program.

An agency decides it wants to use a frontier model on controlled unclassified information. Someone finds a vendor announcement saying the service is FedRAMP High authorized. The requirement gets written, the architecture gets drawn, and the question of whether that specific model is available at that specific level in that specific region gets answered much later, by an assessor.

This is a solvable problem. Both major cloud providers publish the answer, maintained and dated, in tables that almost nobody reads. What follows is those tables read closely, as of September 2026.

Why the announcements mislead without being wrong

Vendor announcements are accurate and incomplete in a consistent way. They name a service and a level, and they omit the region, the deployment cloud, the model, and the date. Each of those four things changes the answer.

Take the clearest example. Azure OpenAI appears in Microsoft’s audit-scope tables twice. In Azure public it is in scope for FedRAMP High and DoD IL2. In Azure Government it is in scope for FedRAMP High, IL2, IL4, IL5WI and IL6. Same service name, two very different scopes, and only the second one reaches controlled unclassified information at impact level 4 or 5.

An agency that reads “Azure OpenAI is FedRAMP High authorized” and deploys in commercial Azure has not made a mistake about the authorization. It has made a mistake about which authorization.

IL5 is not one thing

A second precision problem sits inside the Azure Government column. The heading is not IL5. It is IL5WI, which Microsoft defines as the DoD SRG Impact Level 5 Workload Isolation Provisional Authorization.

Microsoft’s own note is explicit: in the US Gov Arizona, Texas and Virginia regions, services require extra configuration to meet DoD IL5 compute and storage isolation requirements. There is a separate IL5 audit scope for the US DoD Central and US DoD East regions.

So “Azure OpenAI is IL5 authorized” compresses three facts into one claim: it is authorized at IL5 via workload isolation, contingent on configuration the customer implements, in one set of regions, with a different scope in another set. Each of those clauses is a project decision.

4

model families on the Amazon Bedrock table are authorized in US government clouds while carrying no authorization in AWS commercial East and West regions

7 of 15

Anthropic models listed on the Bedrock table carry GovCloud Class D authorization

2 of 10

Meta Llama models listed on the Bedrock table carry GovCloud Class D authorization

3

separate authorization columns on that one table, each resolving to a different list of models

Counts by Qwalora from Amazon Web Services, "Amazon Bedrock models — FedRAMP and DoD CSP SRG (IL4/IL5) certification status," last updated 8 September 2026. See the method note at the end of this piece.

The finding that inverts the assumption

AWS publishes a model-level table for Bedrock with three authorization columns: US East and West at FedRAMP Class C, US GovCloud at FedRAMP Class D, and DoD CSP SRG at IL4 and IL5.

The reasonable expectation is that the government columns hold a subset of the commercial column. Government moves slower, authorization takes time, so commercial gets the model first and government catches up.

That is not what the table shows.

Four model families appear in the government columns and not in the commercial column. OpenAI’s GPT 5.4, GPT 5.6 and the GPT OSS models. Nvidia’s Nemotron Nano and Nemotron Super models. xAI’s Grok 4.3 and 4.6. Google’s Gemma 4.

Four other families are the reverse, present in commercial and absent from both government columns: AI21 Labs, Cohere, Mistral AI and Stability AI.

The two lists are not subset and superset. They are different lists that happen to overlap.

The overlap is narrower than the catalog suggests

Where a family appears in both, the model lists diverge sharply.

Anthropic has fifteen models on the table. The commercial column carries twelve of them, running from Claude 3 Haiku to Claude 4.6 Sonnet. The GovCloud column carries seven, and it is not the first seven: Claude 3 Haiku, 3.5 Sonnet, 3.7 Sonnet, 4.5 Sonnet, 4.8 Opus, 5.0 Opus and 5 Sonnet. Two of those, 4.8 Opus and 5.0 Opus, appear in GovCloud while carrying no commercial authorization at all.

The DoD column is narrower again, dropping both.

Meta shows the same shape more starkly. Ten Llama models in commercial, two in GovCloud and DoD.

Amazon’s own Titan family splits too: four models in commercial, two text-embedding models in the government columns.

What this means in practice

An architecture that assumes model parity across clouds is an architecture that will need revision. A team that prototypes on a commercially available model and plans to lift the workload into GovCloud may find the model is not there. A team that assumes the newest model cannot possibly be authorized yet may be wrong in the other direction.

The only reliable move is to read the table for the specific model, the specific cloud, and the specific impact level, on the day the decision is made.

Four assumptions the published tables do not support

Commonly assumed

What the tables show

A service is FedRAMP authorized or it is not

A service is FedRAMP authorized or it is not

Authorization is scoped per cloud, per region, and for AI platforms, per model

Government cloud carries a subset of commercial

Government cloud carries a subset of commercial

Four Bedrock model families are authorized in government clouds and not in AWS commercial East and West

Newer models reach commercial first

Newer models reach commercial first

Two Anthropic models on the Bedrock table appear in GovCloud with no commercial authorization

If the model is authorized, the surrounding services are too

If the model is authorized, the surrounding services are too

In Azure Government, Azure OpenAI reaches IL6 while Azure AI Content Safety and the Foundry portal stop at IL2

IL5 is a single authorization state

IL5 is a single authorization state

Azure Government publishes IL5WI, workload isolation, requiring customer configuration, with a separate scope for US DoD regions

FedRAMP baselines are Low, Moderate and High

FedRAMP baselines are Low, Moderate and High

AWS now publishes Class C, formerly Moderate, and Class D, formerly High

The guardrails problem

The Azure Government tables contain an architecture constraint that deserves more attention than it gets.

Azure OpenAI is in audit scope for FedRAMP High, IL2, IL4, IL5WI and IL6. Azure AI Content Safety is in scope for FedRAMP High and IL2 only. The Microsoft Foundry portal is likewise FedRAMP High and IL2 only. Microsoft Copilot Studio reaches IL4 but not IL5WI or IL6.

Read that as a system rather than a list. An agency can run the model at IL5. It cannot run Microsoft’s first-party content safety service inside that same boundary, and the portal it would manage the deployment from is not authorized there either.

That is not an argument against building. It is an argument for finding out before the design review rather than after, because the answer changes what the safety architecture has to look like and who builds it.

Some things are explicitly outside the boundary

AWS publishes a fourth column labelled FedRAMP Not Required, and notes that these services are not included in the FedRAMP certification boundary and that customers work with their AWS account team to seek independent agency approval. The AWS Management Console, Cost Explorer, Marketplace and Artifact sit there.

This is worth knowing precisely because it is unremarkable in commercial use. A management plane outside the authorization boundary is a normal condition that has to be documented rather than discovered.

What to do with this

Ask the four questions together. Which cloud, which region, which model, which impact level. Any one of them answered alone produces a claim that sounds settled and is not.

Read the maintained tables, not the announcements. Both vendors publish per-service and, on the AWS side, per-model tables with revision dates. Announcements are accurate on the day they are written and are not revised.

Check the surrounding services, not only the model. The inference service, the safety service, the portal, the orchestration layer and the management plane can each carry a different scope.

Date the finding. These tables move. Anything recorded in a requirements document should carry the date it was read, and should be re-read before the design is frozen.

Design for substitution. If the model you want is not authorized where you need it, that is an architecture question rather than a blocker. An inference layer that abstracts the specific model is worth building on its own merits, and it is worth considerably more when the authorized list changes underneath you.

Method, and what this piece does not establish

Every factual claim above comes from a vendor’s own maintained compliance table, cited below with its revision date. Nothing is drawn from announcement posts, and where an older announcement conflicts with a current table, the table is used.

Three limits worth stating plainly.

The counts in the stat band are ours, derived by reading the AWS model table. They are arithmetic on a published source rather than an independent finding, and anyone can reproduce them.

The AWS model table is dense and the individual cells are easy to misread. Readers making a real decision should open the table and confirm the specific row and column rather than relying on our reading of it.

These tables change without announcement. Everything here was read in September 2026. It should not be treated as current beyond that.

Qwalora holds no partnership, reseller agreement, or commercial relationship with Amazon Web Services, Microsoft, Anthropic, OpenAI, Meta, Google, Nvidia, xAI, Mistral, Cohere, AI21 Labs or Stability AI. This piece describes published authorization status and is not a recommendation of any platform or model.

Sources

  1. Amazon Web Services. Amazon Bedrock models — FedRAMP and DoD CSP SRG (IL4/IL5) certification status. Last updated 8 September 2026. https://aws.amazon.com/compliance/services-in-scope/FedRAMP/amazon-bedrock-models/

  2. Amazon Web Services. AWS Services in Scope by Compliance Program — FedRAMP. Last updated 9 September 2026. https://aws.amazon.com/compliance/services-in-scope/FedRAMP/

  3. Amazon Web Services. Amazon Bedrock in AWS GovCloud (US). AWS GovCloud (US) User Guide. Read September 2026. https://docs.aws.amazon.com/govcloud-us/latest/UserGuide/govcloud-bedrock.html

  4. Microsoft. Azure, Dynamics 365, Microsoft 365, and Power Platform services compliance scope. Microsoft Learn. Service tables last updated February 2026; page revised 18 August 2026. https://learn.microsoft.com/en-us/azure/azure-government/compliance/azure-services-in-fedramp-auditscope

  5. Microsoft. Isolation guidelines for Impact Level 5 workloads. Microsoft Learn. Read September 2026. https://learn.microsoft.com/en-us/azure/azure-government/documentation-government-impact-level-5

Tell us what you are deciding.