Stealth startup's Cortex scores 73.2% on τ³-bench banking, beating frontier LLMs by 18 points
multiply_matrix · x · 2026-11-06
Principled Machines, still in stealth, reports Cortex passed 73.2% of tasks on Sierra's τ³-bench banking benchmark (Sierra-verified) versus 55.2% for the best published model, Qwen 3.8 Max; Claude Opus 5 scored 48.7%, Grok 4.5 47.9%, GPT-5.6-sol 46.9%. Cortex builds algorithmic complements to LLMs rather than fighting what models do well — Sierra's own paper shows even agents given all needed documents passed only 39.7%, so retrieval is half the problem and Cortex targets the other half (acting on information correctly). No task-specific fine-tuning or prompt changes. It will ship as an OpenAI-compatible model API in early access.
More from Companies & People
- Trae Stephens on going from Palantir to Founders Fund to Anduril — adityaag · 2026-11-06
- Founders must hide their fear while financiers flex vulnerability — katieruthmishra · 2026-11-06
- Lambda CEO is extremely bullish on neoclouds as compute demand compounds — gharik · 2026-11-06
- Parallax founding team assembles in UK to build white-box belief auditing for AI evaluators — gsarti_ · 2026-11-06
- OpenAI Sets Nov 12 'Cyber Day' With Livestream and Security Announcements — gdb · 2026-11-06
- Anthropic signs first-ever content licensing deal with the AP for election data — ScottNover · 2026-11-06