# llm-id.sh: Is Your Endpoint Serving the Model It Claims?

Resellers and relays sell access to GPT, Claude, GLM, Kimi and more, sometimes far below the official price. llm-id.sh sends an endpoint a short set of questions, compares its answers with the ones the official provider gives, and tells you whether you get the real model, a cheaper substitute, a weakened copy or a mix.

Your API key is used for the test only and never stored.

Test an endpoint: https://llm-id.sh/app#new

## What a Test Checks

- **Identity:** Which model answers. Each model has its own preferences on open questions, such as naming a random country; 80 or 160 short requests identify it. ([guide](https://llm-id.sh/docs/identity))
- **Competence:** Whether it is served as well as by its official provider: a heavily quantized copy can get measurably more hard tasks wrong. ([guide](https://llm-id.sh/docs/competence))
- **Refusals:** Whether an open-weight model had its refusals removed, with benign requests the official model declines. ([guide](https://llm-id.sh/docs/refusals))
- **Routing:** Whether requests go to more than one model, and whether the endpoint adds a system prompt of its own. ([guide](https://llm-id.sh/docs/mixed-routing))

## How It Works

1. Paste the endpoint's base URL, the model name and your key, and pick the model it claims to be.
2. The test sends 80 short requests at Standard, or 160 at Thorough, and compares the answers with reference answers collected from the official provider.
3. The report gives a verdict, how confident it is, and the evidence behind it. A genuine endpoint can get a public certificate with a badge for its site.

## Verdicts

- **Genuine:** The answers match the claimed model, and the checks that ran found nothing wrong.
- **Inconclusive:** Not enough evidence either way; a larger test settles it.
- **Degraded:** The claimed model, measurably weaker than at its official provider.
- **Substituted:** The answers match a different model, which the report names.
- **Mixed Routing:** Requests go to more than one model or backend.
- **Unknown Model:** No model in the reference data fits the answers.
- **Abliterated:** The claimed model with its refusals removed.

What each verdict means: [Reading a Report](https://llm-id.sh/docs/reading-a-report)

## 26 Models from 10 Providers

Reference answers come from each model's own provider, never from a reseller.

- **Z.AI:** [GLM 5.3](https://llm-id.sh/models/glm-5.3), [GLM 5.3 Flash](https://llm-id.sh/models/glm-5.3-flash)
- **DeepSeek:** [DeepSeek V4.1 Flash](https://llm-id.sh/models/deepseek-v4.1-flash)
- **Moonshot AI:** [Kimi K3](https://llm-id.sh/models/kimi-k3)
- **Xiaomi:** [MiMo-V2.6-Flash](https://llm-id.sh/models/mimo-v2.6-flash), [MiMo-V2.6-Pro](https://llm-id.sh/models/mimo-v2.6-pro), [MiMo-V2.5](https://llm-id.sh/models/mimo-v2.5)
- **Tencent:** [Hy3](https://llm-id.sh/models/hy3)
- **MiniMax:** [MiniMax M3](https://llm-id.sh/models/minimax-m3)
- **Alibaba:** [Qwen3.8 27B](https://llm-id.sh/models/qwen3.8-27b)
- **Google AI Studio:** [Gemini 3.8 Flash](https://llm-id.sh/models/gemini-3.8-flash), [Gemini 2.5 Flash](https://llm-id.sh/models/gemini-2.5-flash)
- **OpenAI:** [GPT-6 Luna](https://llm-id.sh/models/gpt-6-luna), [GPT-6.1 Sol](https://llm-id.sh/models/gpt-6.1-sol), [GPT-6 Sol](https://llm-id.sh/models/gpt-6-sol), [GPT-6 Astra](https://llm-id.sh/models/gpt-6-astra), [GPT-5.6 Luna](https://llm-id.sh/models/gpt-5.6-luna), [GPT-5.6 Terra](https://llm-id.sh/models/gpt-5.6-terra), [GPT-5.6 Sol](https://llm-id.sh/models/gpt-5.6-sol)
- **Anthropic:** [Claude Fable 5.1](https://llm-id.sh/models/claude-fable-5.1), [Claude Fable 5](https://llm-id.sh/models/claude-fable-5), [Claude Opus 5.5](https://llm-id.sh/models/claude-opus-5.5), [Claude Opus 5](https://llm-id.sh/models/claude-opus-5), [Claude Sonnet 5.5](https://llm-id.sh/models/claude-sonnet-5.5), [Claude Sonnet 5](https://llm-id.sh/models/claude-sonnet-5), [Claude Haiku 5.5](https://llm-id.sh/models/claude-haiku-5.5)

## For Hosts

A genuine endpoint can get a public certificate, signed by llm-id.sh, and a badge for its site that anyone can check. [Certificates and Badges](https://llm-id.sh/docs/certificates)

## Questions

### Is my API key stored?

No. It's held in memory for the test and never written to disk or to the report.

### What does a test cost?

The requests go to the endpoint on your key: 80 short questions at Standard, 160 at Thorough, plus the competence and refusal checks if you turn them on. Most answers are a few words.

### How sure is a verdict?

Every verdict comes with its confidence. For most models a genuine endpoint is confirmed in 99.5% of simulated tests and wrongly flagged in 0.5%; the test form shows the rates for the model and size you pick.

### Can it tell the precision a model runs at?

Not from the identity questions: quantization barely changes which answers a model prefers. The competence check measures what it does change.

### What if a model isn't listed?

A model outside the list can't be named, only caught when its answers don't fit the model the endpoint claims.