Answers from our docs and articlesesc to close
Try asking
Create an accountSign up

Which model providers don't train on your prompts, and which keep nothing

Of 598 certified provider routes, 432 are listed as not training on prompts and 383 as zero data retention. Where they are, and how to require them.

Key takeaways

  • On Oct 22, 2026, 432 of Inference Bureau's 598 certified routes were listed as not training on prompts, and 383 as zero data retention.
  • Zero data retention is the stricter guarantee: every route listed with it was also listed as not training, while 49 routes were no-training only.
  • The same model can qualify on one provider and not another, so the guarantee belongs to the route, not the model.
  • Set data.requireZdr or data.requireNoTraining in Directives, and every request goes only to routes that qualify, with the upstream gateway told to refuse the rest.

When you send a prompt to a model, the provider that runs it decides what happens to it next. Some keep nothing once the response is sent. Some keep request data for a while, for abuse monitoring or debugging, but never train on it. Others make neither promise. The model's name tells you nothing about which applies, because the same open model is often served by a dozen providers with different terms.

Inference Bureau records two flags for every certified route, meaning one model on one provider, and routes on them when your Directives ask it to.

What the two flags mean

Both flags come from the public endpoint listing of the upstream gateway Inference Bureau sends requests through. For each model, the listing gives every provider endpoint, and for each endpoint whether it has zero data retention and whether it trains on request content.

  • No training (noTraining): the provider endpoint is listed as not training on request content. It may still keep requests for a time.
  • Zero data retention (zdr): the provider endpoint is listed as keeping no request data. It is the stricter of the two. It also rules out providers that store data for a while without training on it.

These are the provider's terms as the gateway lists them. Inference Bureau does not audit providers or test what they keep. What it does is route only to the endpoints that qualify, and tell the gateway, on every request, to refuse any endpoint that doesn't.

Inference Bureau keeps no prompt or response content of its own by default. It is stored only when you turn on the Log content Directive, and never where zero data retention applies.

The numbers, by provider

The certified catalog read on Oct 22, 2026 held 598 routes: 274 models across 59 providers.

FlagRoutesShare
Zero data retention (and no training)38364%
No training only498%
Neither16628%

No route was listed as zero data retention without also being listed as not training. So requiring ZDR is a strict subset of requiring no training.

Most providers are all or nothing. These are the providers with the most routes:

ProviderRoutesZDRNo training
openai555555
deepinfra503838
bedrock494048
novita4500
alibaba362525
azure333333
vertex262626
anthropic171517
google17017
gmicloud1500
zai1400
baseten141414

Some of the splits are worth knowing:

  • Gemini on Google's own API vs Vertex. All 17 routes on google were listed as not training, and none as zero data retention. The 26 routes on vertex had both.
  • Anthropic. 15 of 17 routes on Anthropic's own API had both flags. The other two, claude-fable-5 and claude-fable-5.1, were no-training only there and on Bedrock and Vertex too, so no provider offered them with ZDR.
  • Bedrock. 40 of 49 routes had ZDR. The nine without it were six OpenAI models, the two Fable models and claude-sonnet-4. OpenAI's models had ZDR on openai and azure.
  • Model makers' own APIs. Several labs' first-party endpoints had neither flag on any route: zai (14 routes), minimax (9), deepseek (4), xiaomi (5) and tencent (5). Their models are often available with ZDR from other providers.

The numbers, by model family

A model qualifies when at least one of its certified providers does. 221 of the 274 models had at least one no-training route, and 204 had at least one ZDR route. For the vendors with the most models:

VendorModelsWith a ZDR routeWith a no-training route
openai646464
alibaba322525
google191819
anthropic181517
zai1799
meta131111
deepseek121111
minimax933
mistral999
moonshotai866

The open-model families are where it matters which provider you land on. Take zai: of its 17 models, 9 had a ZDR route, and those 9 spread 40 ZDR routes across other providers, while Z.ai's own endpoint had none. 53 models, including moonshotai/kimi-k2 and most minimax models, had no provider listed as not training at all.

Requiring it with Directives

Set the rule in the Data group of the Directives table, at the team, a project, an environment or a key. In JSON it is the data field:

{
  "data": { "requireNoTraining": true }
}

For zero data retention, set requireZdr instead. Set both to require both, though in today's catalog ZDR already implies no training:

{
  "data": { "requireZdr": true, "requireNoTraining": true }
}

With the management API, change one field and keep the rest:

curl -X PATCH https://api.inferencebureau.com/v1/projects/inference-bureau/directives \
  -H "Authorization: Bearer $IB_MANAGEMENT_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "policy_version": 3, "set": { "data.requireZdr": true } }'

Here is what follows:

  • Set at any level, the rule applies to everything below it and can't be lifted there. A team that requires ZDR covers every project and key.
  • Providers that don't qualify are skipped, with no_zdr or no_training on the request's receipt in Logs. A model with no qualifying provider returns 403 policy_denied and is left out of GET /v1/models.
  • Every request also carries the gateway's own flags, which refuse a provider that doesn't qualify, so a change of terms is caught upstream as well.
  • Under requireZdr, nothing is stored on Inference Bureau's side. Stored responses on /v1/responses are refused, and turning the rule on deletes those already stored under that level.
  • Under either rule, /v1/messages/count_tokens answers with a local estimate rather than sending the prompt to a provider.

How the flags stay current

Providers change their terms. The catalog records each route's flags when it is synced, and the counts above are from that record. Between syncs, a scheduled job at 37 minutes past every hour reads the gateway's listing for every certified model. It needs no credential and runs no inference. It stores the result as a snapshot that every process planning requests reads.

The snapshot can only take a guarantee away. If the gateway stops listing a provider as ZDR or as not training, that route stops qualifying within the hour. A newly listed provider is used only once Inference Bureau records it in the catalog. A snapshot older than 48 hours is ignored, and so is one that looks like a gateway fault: one that couldn't read more than a fifth of the models, or that would revoke more than half the guarantees at once. Planning then falls back to the catalog's own flags. The per-request flags still apply either way.

Frequently asked questions

Does Inference Bureau verify that a provider keeps nothing?

No. The flags are the provider's terms as the upstream listing gives them. Inference Bureau routes on them and asks the gateway to refuse any provider that doesn't qualify, but it does not audit providers.

Is zero data retention the same as not training on prompts?

No. No training means the provider doesn't train on request content but may keep it for a time. Zero data retention means it keeps none. In the Oct 22, 2026 catalog, every ZDR route was also listed as not training.

What happens if a model has no qualifying provider?

Requests for it return 403 policy_denied, and it is left out of GET /v1/models for keys under the rule. The model isn't swapped for another unless you set up model fallback.

Does Inference Bureau store my prompts?

Not by default. Content is kept only when the Log content Directive is on, for 7 or 30 days, and never where requireZdr applies.

One bill for your apps, your models and your reviews.

Host apps and functions next to your model traffic. Routing picks each turn's model by your quality floor, and Auditor reviews every pull request against your own rules.