AI provider transparency

All rankings measure how smart models are. This one measures something different: whether the provider takes you seriously. Whether they post pricing without a sales call, admit to outages, document what changed in the model, and let you run anything on your own infrastructure. Seven criteria, each verifiable by clicking a link, with no room for my personal preferences.

last scan: 2026-11-09 06:41:24 · maximum 7 points · green cell links to proof

What exactly I check

• Public pricing - public pricing page, no form and no "contact us for quote". First test of respect for the customer.
• Status page - a place where the provider admits to outages before your customer does. Without it, you'll learn about problems from social media.
• Changelog - list of changes to models and API. Without it, a model can start responding differently overnight and nobody will warn you.
• API documentation - public, no login or corporate account required.
• Data policy - clear document about what happens to your prompts and whether they go into training.
• Model card - description of limitations and risks, not just a list of benefits from press materials.
• Open weights - whether the provider publishes any models to run on your own infrastructure. It doesn't have to be the flagship: what matters is whether they release weights at all, because that's the only emergency exit when they change pricing or shut down the service.

Methodology, including one thing I don't hide

Checking presence and availability of the resource at the address, not its quality. A provider may have a changelog and write generalities in it - the ranking won't catch that, but it will catch complete absence. I manually verified all 42 addresses on July 31, 2026, and automation checks them weekly.

Technical detail worth knowing, because it disqualifies most similar comparisons: some services (including OpenAI and xAI) reject requests from data center servers with a 403 code. A naive scanner would treat this as "page doesn't exist" and give them 2 out of 7 instead of the real score. My scanner distinguishes three responses: confirmed "exists", confirmed "doesn't exist" (404 or no domain), and "don't know" - and with "don't know"leaves the last confirmed state instead of making up a result. Two real absences detected during verification: xAI has no changelog at the address from documentation (404), and Moonshot AI has no public status page (domain doesn't respond).

When citing, provide "quantslant.com, transparency ranking" and scan date. I have pricing for all models in cost calculator, current service availability in status monitor, and the price per unit of capability in price of intelligence.