Request Budget

Who handles your request?

No training, no stored content and regional processing are three different promises.

Request-path sketch only. Position and line length do not show speed or security.

Model maker, host or gateway?

Request-path sketch only. Position and line length do not show speed or security.

Start with the retention switch

OpenAI

Default API handling
No training without opt-in; abuse logs up to 30 days

OpenAI API

Legal/safety exceptions can extend retention. ZDR requires approval; endpoint and feature limitations may still retain application state. Responses default application state is a separate control.

Source ↗

Together AI

Organization privacy controls
Prompt storage on; model training opt-in off

Together organization privacy settings

Turning prompt storage off enables ZDR and disables passthrough models. Passthrough sends data to upstream providers under their own policy.

Source ↗

Groq

Inference retention
No customer-content retention by default; exceptions apply

GroqCloud inference

Persistence features and reliability/abuse troubleshooting can retain content. Logs up to 30 days unless law requires longer; Data Controls opt-out. Usage metadata retained.

Source ↗

OpenRouter

Content retention
Opt-in and safety/legal exceptions

OpenRouter prompt/response logging

Logging/product improvement off by default. Safety Retention endpoints can retain any request for abuse monitoring; exclude them with ZDR privacy settings. CSAM/legal exceptions and metadata remain.

Source ↗

Then check the selected model or feature

Anthropic

Fable / Mythos 5.2
30-day retention required unless expressly authorized

Claude Fable 5.2 and Claude Mythos 5.2

Not available under ZDR unless Anthropic expressly authorizes it. Do not generalize to every Claude model.

Source ↗

Google Vertex AI

Grounding retention
Search up to 3 days; Maps 30 days

Grounding with Google Search and Grounding with Google Maps

Search logs include derived queries/context; Maps stores prompts/context/output. Cannot disable this storage while using those features. Advanced AI features may have additional logging; no blanket ZDR claim.

Source ↗

Google Vertex AI

Interactions API
store defaults to true

Managed models through Interactions API

Explicit store=false needed for that application state. Request-response BigQuery logging is separately off by default; other retention features still need review.

Source ↗

Amazon Bedrock

Retention mode
none blocks a model that requires retention

Account per Region; project setting only on bedrock-mantle

New scope inherits model default. store=false alone is not ZDR. none prohibits durable request/response storage; Responses store=true rejected and background mode unavailable.

Source ↗

Amazon Bedrock

Fable 5 / 5.1 / 5.2 on Bedrock
AWS may retain and review content for up to 30 days

Models requiring aws_review, currently Claude Fable 5, 5.1, 5.2

Retained within the AWS boundary; not shared with model provider. Cross-region inference stores retained inputs/outputs in destination processing Regions. ZDR exceptions require per-account/per-model eligibility.

Source ↗

Processing location is only one boundary

Microsoft Foundry

Processing scope
Global / Data Zone / Azure geography

Foundry Serverless API deployment types, not managed compute

At-rest data stays in designated Azure geography. Global can process any Azure region; Data Zone within US/EU/APAC zone; Standard/Regional Provisioned within geography, possibly between its regions.

Source ↗

Together AI

US processing endpoint
Select Serverless models can process in the US

us-api.together.ai, select Serverless models

This processing scope does not guarantee storage, logging or routing location. Other residency requirements may require dedicated/VPC deployment.

Source ↗

OpenRouter

In-region routing
Business and Enterprise; US/EU

OpenRouter US/EU in-region routing

Standard has data-policy routing but not these in-region features. Contractual SLA Enterprise only; paid rate limits pass through underlying provider.

Source ↗