GPT-6 Astra sets ECI record, surges on science benchmark

Epoch AI, granted pre-release access by OpenAI, has evaluated GPT-6 Astra: Astra set a new record of 169 on the Epoch Capabilities Index (ECI), surpassing the previous best of 163, and also set new records across multiple benchmarks including math, continual learning, and game-based puzzles. On Terminal-Bench-Science 0.1, Astra took the top spot with 64.6%, up 42.2 percentage points from GPT-5.6 Sol's 22.4%. Launched just a week earlier, the benchmark became the headline showcase at the GPT-6 Astra launch event, with third-party quantitative data offering a reference point for gauging this generational leap.

Confirmed

• After testing GPT-6 Astra, Epoch AI reported an ECI score of 169, breaking the previous record of 163
• Astra set new records on multiple benchmarks including math, continual learning, and game-based puzzles
• Terminal-Bench-Science 0.1 launched a week ago: GPT-5.6 Sol scored 22.4%, GPT-6 Astra reached 64.6%, a 42.2 percentage point gain

Why it matters

• While the ECI jump is notable, it reportedly still falls within the uncertainty band of Epoch AI's reasoning-era ECI trend, suggesting it remains to be seen whether the capability gain exceeds the existing trend
• The rapid rise of a new benchmark to centerpiece of the launch reflects how benchmark selection shapes model marketing

2026-09-04 ~ 2026-09-04 · 6 related posts

• Epoch AI: GPT-6 Astra sets ECI record of 169, tops math and continual learning benchmarks — NathanpmYoung · 2026-09-04
• [source] Epoch AI publishes GPT Astra results on its ECI benchmark — Wonderful_Buffalo_32 · 2026-09-04
• [source] GPT-6 Astra jumps 42 points to 64.6% on Terminal-Bench-Science, launched last week — btibor91 · 2026-09-04
• GPT-6 Astra sets new records on Epoch AI's Capabilities Index, math and learning benchmarks — JosephJacks_ · 2026-09-04

2 near-duplicate retellings: BenBlaiszik · Jsevillamol