In one sentence
Before verification, our index gave 31 out of 100 to 37 of 38 proprietary models. After reading the official documentation, they range from 44 to 80. The first number measured our ignorance. The second measures what vendors publish.
Why a barometer
A CIO choosing an AI model has to answer simple questions: where does my data go, is it used for training, can I have it deleted, do I get a DPA, SSO, audit logs. The answers exist. They are scattered across terms pages, trust centers and technical documentation, and they change without notice.
The nAIvigate Trust Index reads those pages line by line and scores what the vendor documents. Twenty criteria, five families, one rule: what is not documented counts as zero and stays visible. It is not a compliance score, compliance depends on your processing. It measures what the vendor offers and publishes.
This is the first edition. It will be updated at every significant change, and at least monthly.
What we did on 9 September
We verified the commitments of 12 vendors covering 38 proprietary models, on their official pages only: platform documentation, DPA, trust center, subprocessor list, API terms. For 33 models the verification is dated and sourced. For 5, we found no usable page, and the index says so.
One scoping rule became unavoidable along the way: we score the vendor's first-party platform, not what it offers through a hyperscaler. Claude on Bedrock in Frankfurt is not Claude on api.anthropic.com. A customer going through AWS is covered by AWS's DPA. Mixing the two means scoring the host instead of the vendor.
The results
| Vendor | Platform scored | Score | Documented | |---|---|---|---| | Google | Vertex AI | 80 | 100% | | OpenAI | API Platform | 64 | 85% | | xAI | x.ai API | 55 | 85% | | Anthropic | Claude API | 51 | 85% | | Alibaba | Model Studio | 45 | 70% | | Black Forest Labs | FLUX API | 44 | 55% |
Models from the same vendor share their platform's score, unless an exception is documented: DALL-E 3 is excluded from Zero Data Retention at OpenAI, Claude Fable 5 enforces 30-day retention at Anthropic, Sora 2 is hosted in the US only.
Three things the table reveals
Anthropic offers no European region on its direct API. The official documentation lists two inference geographies, us and global, and a single storage geography, us. EU-region access goes through Bedrock or Vertex, outside the vendor's scope. Our database said "US/EU" by inheritance. It now says "US", and Claude loses two of five criteria in the Data family. It is the most visible correction of this edition, and the best documented.
The top-ranked European vendor uses your data for training, unless you opt out. Black Forest Labs, the German maker of the FLUX models, offers a European endpoint. Its API terms grant the company a licence on your inputs and outputs, including for training its models. Opting out is done by email. An EU endpoint is not a processing guarantee.
Google leads, but only through Vertex AI. The six Gemini models reach 80 with 100% of criteria documented: EU regions per model, no training, Zero Data Retention, SSO, SCIM, audit logs, private endpoint, published subprocessors. The Gemini Developer API does not offer these guarantees. The score says "Vertex", not "Gemini".
What remains undocumented
Even among the best scored, three criteria are missing from public pages for the API: SSO/SAML, SCIM provisioning and private endpoint. OpenAI, Anthropic and xAI offer them in their Enterprise or chat products, but do not document them for programmatic access. They count as zero in the index until published. If a vendor sends us the page, the score rises the same day.
Method and limits
Score = weighted average of five families: data 30%, sovereignty 20%, enterprise 20%, transparency 15%, licence 15%. A criterion is worth 100, 0, an intermediate value (on request, depends on plan, opt-out) or 0 if undocumented. Below 50% of criteria documented, the score is marked provisional and does not rank. No position is for sale. The full method, the weights and the source of every value are on the Trust Index page and on each model page.
The limits are those of the method: we read what is published, we do not test the commitments, and a page can change between two verifications. Every model page carries its date. Each model's trust profile can be downloaded as a PDF from its page.
What it changes for you
If you run one of these models in production, the question is no longer "is it compliant" but "on which platform, with which dated commitments". And if you do not know which ones your teams already use, that is where to start: it is the first deliverable of nAIvigate Studio's AI Radar.
nAIvigate will be at Big Data & AI Paris on 15 and 16 September, Porte de Versailles.