01Oct 09, 2026
Before checking whether a third-party model API behaves like its advertised model, configure the same endpoint, model ID and API protocol you use in your application. A successful connection alone does not establish model authenticity.
1. Base URL — Copy the API base from your provider's integration documentation, including any required version prefix. Do not paste the dashboard URL or a full chat/completions or messages request URL. For example, if the provider documents a base ending in /v1, keep that prefix; do not add /v1 blindly.
Read article →
02Oct 09, 2026
An HTTP error describes how the endpoint handled a request; by itself, it does not show whether the advertised model is genuine. Providers and gateways can use status codes differently, so check the response and the provider's own API documentation together.
401 usually points to missing, invalid, expired, or incorrectly scoped credentials. Confirm that the key belongs to this provider, is sent in the format the API expects, and can access the requested model. Do not paste the key into logs or support posts.
Read article →
03Oct 09, 2026
Use the endpoint and model identifier that your application sends to the provider. Dashboard pages, model marketing names, and API request values can be different.
Find the Base URL in the provider's API or integration documentation. Copy the documented API root, including a required version prefix such as /v1. Do not use the provider dashboard URL, and do not include the full operation path or query string. Providers can use different base paths, so do not add /v1 unless their docs specify it.
Read article →
04Oct 19, 2026
AI model verification compares a third-party API's observed behavior with a selected official reference. It provides evidence about model identity, not cryptographic proof of the model running behind a gateway.
Model verification compares endpoint behavior against official model fingerprints, while benchmark scores serve as a reference for capability comparison rather than a promise about any one provider.
Read article →
05Oct 25, 2026
Model verification checks whether a third-party endpoint behaves like its claimed model by comparing against official fingerprints — without disclosing protected verification assets.
The workflow sends diagnostic probes to the target endpoint, then compares the response pattern to a known-good fingerprint collected from genuine official endpoints.
Read article →
06Jun 07, 2026
Public leaderboard scores are grouped into strong, moderate, and weak bands so readers can compare relative capability quickly.
The baseline_150 board is a reference layer: it helps you compare official model behavior and gives context when a verification run claims to match a specific model.
Read article →
07Oct 25, 2026
Many third-party AI services advertise OpenAI- or Anthropic-style APIs, but protocol compatibility does not guarantee that the endpoint runs the model it claims to run.
Start with the exact API base URL and request model ID used by your application. Keep a version prefix such as /v1 only when the provider documents it. Use the API root rather than a dashboard URL or a complete chat/completions route.
Read article →
08Oct 19, 2026
To investigate whether a GPT API serves its advertised model, distinguish connection problems from behavioral differences. A returned model name, fluent answer or OpenAI-compatible response format does not independently prove model identity.
Confirm the documented base URL, exact model ID, key permissions and intended official reference first. An HTTP error or handshake failure means the check could not complete; latency alone cannot establish that a model has been substituted.
Read article →
09Jun 07, 2026
Your API key is stored transiently in Redis, used only for the current run, and deleted immediately after the run completes. It is never written to PostgreSQL or the file store.
The frontend never expects or displays decrypted keys after save.
Read article →
10Jun 08, 2026
Model mismatch risk is the practical downside of buying access through a third-party endpoint without independent evidence about what model is really behind it.
The direct cost is overpaying for weaker capability. The indirect cost is harder to see: bad routing decisions, misleading evaluations, latency regressions, and support incidents caused by behavior you did not expect.
Read article →
11Jun 08, 2026
The public benchmark methodology is intentionally partial: readers can see what the score is meant to represent without getting access to protected prompts, grading rules, or answer keys.
In practice, that means publishing category-level outcomes and leaderboard positions while keeping the exact evaluation assets private enough to preserve the usefulness of future comparisons.
Read article →
12Jun 07, 2026
Public pages expose official benchmark summaries, while anonymous test reports remain accessible only by saved report code until they expire.
This lets the product support shareable evidence and follow-up analysis without requiring a permanent account dashboard for every public user.
Read article →
13Jun 07, 2026
What is model verification? | Model verification checks whether a third-party API endpoint behaves like its claimed model by comparing against official fingerprints — without disclosing protected verification assets.
How does ModelVerify.ai check model identity? | The platform runs diagnostic probes against the endpoint and compares the responses to official model fingerprints collected from known-genuine endpoints. Verdicts should be interpreted alongside latency, stability, and drift signals.
Read article →
Methodology overview
Public docs explain methodology and report interpretation without exposing protected prompts, probes, official outputs, or private thresholds.