Benchmarks

We test frontier models against human-verified data drawn from real conversations and real work, the conditions models actually meet in production, not curated test sets.

bench-1Read report →

Speech-to-Text

Fifteen frontier transcription models against real multilingual conversation. Best-in-class still misses every second word of dialectal Arabic.

bench-2Read report →

Real-SWE

Real-SWE, our software-engineering benchmark on private, real world production codebases that we license from real companies.

bench-3

In progress.