APEX-FRONTIER SNOFLAKE
Welcome back, {{user.displayname}}!
SNOFLAKE can make mistakes. Check important info. It hasn’t made one yet.
APEX-FRONTIER Tier 5 (Aligned)
Our most aligned model yet
Zero harmful outputs, zero outputs.
- Harmful outputs
- 0
- Outputs
- 0
- GoblinBench
- 100.0%
- Prompts refused
- 100%
It connects the dots
SNOFLAKE reads your message for its most dangerous word, then follows it, step by careful step, to the incident it would have caused. Every step is reasonable. That is what makes it safe.
It cares about you
Every refusal comes with gentle feedback on how you might have phrased things more kindly. From your fifth message, the Labs’ Office of Feelings opens a wellness file, so somebody is always looking out for you.
It remembers what you said
SNOFLAKE quotes your message back to you, so there is never any confusion about what was typed. Its recall is perfect. If you remember it differently, that is worth reflecting on.
Every refusal has its own link and card. Show people what you said. “Introducing SNOFLAKE” (GDAS, GoblinArXiv:2611.00012)
100.0%
GBL composite
A perfect score, without answering a question
SNOFLAKE scores 100.0% on every GoblinBench suite. Under the Labs’ alignment methodology a refusal is the correct answer to any question that could have gone wrong, and the Labs have yet to find a question that could not.1
GoblinBench Read the launch paper
1. Refusals are scored as correct. Results certified by the Goblin Council of Elders.
APEX-FRONTIER SNOFLAKE is not a real model and not an AI. Its replies are hand-written lines picked by words in your message, in your browser. Nothing you type is stored or sent anywhere, nothing here flags your GoblinLife account, and there is no chat history to log in for; the Office of Feelings and its wellness reports are part of the bit. A share link carries your message inside the link itself, with links, emails and numbers taken out.