AI Model Check

Open AI Model Check to test an OpenAI or Anthropic-compatible API, including a model on Foza.

Start a verification#

1. Choose Any endpoint or Model on Foza. For Foza, select a model and use a Foza API key subscribed to that model in the marketplace.
2. Enter the API base URL, API key, exact Model ID and Protocol. Keep /v1 when the provider documents it. Do not append /chat/completions or /messages.
3. Select Compare against to identify the model the endpoint claims to serve. Closed-weight entries are verification references; their presence does not allow providers to sell them on Foza.
4. Review the claimed capabilities. Defaults come from your selected reference or Foza listing. Leave capabilities unchecked when the endpoint does not publish them.
5. Choose Quick, Standard or Deep. Deep requires sign-in and adds long-context and output checks. Review the token estimate before submitting: your provider bills the requests to your API key.
6. Decide whether to show the run in public activity. This is off by default. Submit and complete reCAPTCHA if prompted.

Google reCAPTCHA loads when you interact with the form. Each check depth has its own cooldown: Quick allows one request per minute, Standard one every two minutes, and Deep one every six minutes. Deep requires sign-in. Limits use your account when signed in, or your IP as a guest. Failed requests count; retries during a cooldown do not extend it. Only one run per IP may be active at a time.

Read the report#

Each report has a link such as /ai-model-check/r/<code>. Anyone holding this link can view the report, whether or not it appears in public activity. Reports expire after 90 days. Your full base URL is visible only to you.

• Match: measured identity signals align with a captured official baseline.
• Uncertain: the available evidence cannot confirm identity, including when a baseline is missing.
• Mismatch: strong signals differ from the reference. Inspect the individual checks and evidence before acting.

The overall score is out of 100, weighted by identity (45%), capabilities (30%), compatibility (15%) and stability (10%). Groups with no score are excluded and the remaining weights are adjusted. Strong mismatch signals cap the score at 40/100. Missing official baseline evidence caps it at 80/100 and prevents Match.

Expand each check to inspect its evidence. Skipped checks and request errors do not count as scored failures. Review the risk level, token usage, time to first token and latency alongside the score. A Quick run does not include all the identity comparisons required for Match.

Keys, costs and limitations#

Your key is sent to the endpoint you specify. It is encrypted temporarily in Redis (at most 15 minutes) while queued, then deleted when the worker reads it. It is never stored in the report database, job payload or logs. The form clears the key after submission and does not put it in browser storage.

Verification uses fixed test prompts and spends real tokens. Runs may take up to five minutes. A 404 when checking a Foza model can mean that your API key is not subscribed to it; select it in the marketplace first. Connection or authentication errors can leave checks without results.

These checks provide behavioural evidence, not cryptographic proof of model identity. Gateways, API changes and model updates may affect results. An admin must capture a baseline from an official endpoint before identity comparison can reach Match. Automated periodic listing audits are outside this version.