Cortex hits 73.2% on τ³-bench banking, 18 points above best published model
multiply_matrix · x · 2026-11-10
Principled Machines reports Cortex scored 73.2% on Sierra's τ³-bench banking benchmark vs 55.2% for the best published model (Qwen 3.8 Max), with Claude Opus 5 at 48.7% and GPT-5.6-sol at 46.9% — no task-specific fine-tuning or prompt changes. Key insight from Sierra's own paper: agents given all needed documents still passed only 39.69%, so retrieval is half the problem; Cortex builds "algorithmic complements" for the other half. Early access via an OpenAI-compatible API.
More from coding & agent
- NODES session: Test-time strategies like confidence decoding and schema filtering for reliable Text2Cypher — JeremyCMorgan · 2026-11-10
- Stacking More MCP Servers Made My Agent Worse, Reddit User Says — Glittering-Bus-8500 · 2026-11-10
- Anthropic engineer's AI-built browser FPS nears COD quality, sparking 'one year to Sea of Thieves' claim — ChrisGPT · 2026-11-10
- Tech debt in the AI era: paid in full whenever the next frontier model ships — doodlestein · 2026-11-10
- Claude-coded Valheim mod rebuilds Ocarina of Time's Kokiri Forest from your own ROM — Any-Explanation3814 · 2026-11-10
- App Store Connect CLI 6.5 adds subscription retention messages and offers — rudrank · 2026-11-10