Skip to content

Model catalog

Which models vote on your claims

Stamp AI, the default check, rules with one fixed, fast voter of our own choosing. The Research Panel is the enterprise-grade depth: up to 4 models from different providers vote on every claim and a separate judge weighs their reasoning into one verdict. These are the models you can put on that panel. Each adds its credit weight to the per-claim price, so the quote reflects the panel you chose.

ModelProviderCredit weight / claim
Claude Opus 5Anthropic8
Claude Sonnet 5Anthropic5Recommended panel
GPT-6 AstraOpenAI17
GPT-5.6 SolOpenAI4
GPT-5.4 MiniOpenAI2Recommended panel
Gemini 3.8 FlashGoogle3Recommended panel
Gemini 3.1 ProGoogle4
DeepSeek V4.1 FlashDeepSeek1Recommended panel
Grok 4.6xAI2
Kimi K3Moonshot AI5
GLM 5.3Z.ai2
GLM 5.3 FlashZ.ai1
Qwen3.8 2.4TQwen2
Llama 4 MaverickMeta1
Mistral Medium 3.5Mistral3

We add and retire models as providers ship them. This page reads from the live catalog, so it is always current.

How the panel works

  • Panel votes. Each model on your panel researches every claim on the live web on its own and cites its sources. (API callers can also run the panel from model knowledge alone, Model Fact-Check.)
  • A separate judge. Claude Opus 5 reads the panel's votes, reasoning, and evidence and writes one consensus verdict with a calibrated confidence score. Organization admins can pick a different judge.
  • An adversarial challenger. When the panel is unanimous and confident, a cheap model that is not on the panel tries to refute the verdict before we confirm it. If it finds a dissent, the claim goes to the full judge.
  • Stamp AI and Spell & Grammar run on fixed models and aren't part of the picker.

Read the full pipeline write-up on the methodology page, or see pricing for how credit weights turn into a quote.

Customize your panel

Signed-in organization admins choose which models vote and which judge decides on the panel settings page. New accounts start on the recommended panel above.