GPT 5.6 Terra Praised for High Cost-Performance

scaling01 · x · 2026-07-11

Reports claim GPT 5.6 Terra (high) scores 78.3% on WeirdML, closely matching Opus 4.6 (high) but at an allegedly order-of-magnitude lower cost. It also reportedly uses only about 3.5k output tokens per call in high mode and generates shorter code. The commenter found this "a bit off," implying its behavior differs from other GPT models.

Original post →

More from Models

Models channel →