Telluvian vs Patronus AI: hallucination detection compared
Telluvian and Patronus AI both detect hallucinations. Telluvian scores every token in real time by reading a proxy model's internal activations, with no source document needed. Patronus AI offers evaluators that judge a response with another model, billed per call; its Lynx model, for example, returns PASS or FAIL on whether an answer is faithful to a document you supply.
Last reviewed
At a glance
| Telluvian | Patronus AI | |
|---|---|---|
| Model selection method | Prompt-aware. Model Select predicts each model’s performance on the specific prompt and picks the cheapest one that clears your xPerf quality bar. | Not a model router. It evaluates the outputs of models you call yourself. |
| Models and providers | 190+ models from OpenAI, Anthropic, Google, Qwen, DeepSeek, xAI, Z-AI, Moonshot and Meta, behind one OpenAI-compatible API. | Not applicable. An evaluation platform, not a gateway to models. |
| Pricing model | Prepaid, pay as you go. Provider list price for tokens, plus $0.05 per 1M input tokens for Model Select and $1.00 per 1M completion tokens for hallucination scores (optional). | $10 of free credit to start. $10 per 1,000 small evaluator calls, $20 per 1,000 large evaluator calls and $10 per 1,000 explanations. Enterprise is custom. |
| Hallucination detection | Built in. Per-token hallucination scores from probes reading an open-weight proxy model, including for closed models. Off by default. | Yes. Evaluators through its API. Its Lynx model checks whether an answer is faithful to a question and source document and returns PASS or FAIL with reasoning. |
Key differences
Method
Telluvian reads a proxy model's internal activations; Patronus evaluators use a model to judge the response.
Granularity
Telluvian scores each token; Lynx returns one verdict per answer, with reasoning.
Reference text
Telluvian's probe needs none; Lynx checks an answer against a supplied document.
Pricing
Telluvian charges per completion token; Patronus charges per evaluator call ($10 per 1,000 small, $20 per 1,000 large).
Frequently asked questions
What is Patronus Lynx?
Lynx is a hallucination evaluation model from Patronus AI, fine-tuned from Llama 3 70B Instruct. Given a question, a document and an answer, it returns PASS or FAIL with bullet-point reasoning on whether the answer is faithful to the document.
Can I use Lynx commercially?
The published Lynx weights are under a Creative Commons BY-NC 4.0 licence, which restricts commercial use. Check with Patronus AI for commercial terms.
How does pricing compare?
Patronus charges per evaluator call: $10 per 1,000 small evaluator calls and $20 per 1,000 large ones. Telluvian charges $1.00 per 1M completion tokens for hallucination scores, so cost scales with how much the model writes, not with how many checks you run.
Does Telluvian need a source document to detect hallucinations?
No. The probe reads the internal state of an open-weight proxy model as the response is replayed through it, so it can score answers with no reference text. Source-grounded checks such as Lynx answer a different question: does this answer match this document?
Can I see which part of the answer is wrong?
With Telluvian, yes: scores come back per token, alongside the decoded tokens, so you can highlight the exact words that look invented. Lynx returns a verdict for the whole answer, with written reasoning.
Try it on your own prompts
Point an OpenAI SDK at Telluvian, send telluvian/gallery-1 as the model, and see which model answers each request. Questions about your use case go straight to the team.
Sources
Competitor details come from their own public docs and pricing pages, checked on . Where a page did not say, neither do we. Spotted something out of date? Tell us at hello@telluvian.ai.