SecondSourceJudgment rebuilt from primary sources

Epoch AI, an independent research group that runs its own benchmarks on AI models, found that GPT-6.2 Astra can chain 12.3 steps internally without writing out its reasoning, against 10.5 steps for the previous GPT-6 Astra.

Concept illustration: under a black banner lettered EPOCH AI, a framed beige panel shows two matching scenes side by side, each with a laptop whose dark screen shows a grey speech bubble holding a white line, a notebook with a pen lying on it, and a staircase rising to the right from the same ground line. On the left the staircase is white, above a black label reading GPT-6 ASTRA; on the right it is red and taller, above a black label reading GPT-6.2 ASTRA. Printed on the image: 12.3 steps Steps GPT-6.2 Astra chains without writing out reasoning, per Epoch AI, vs 10.5 for GPT-6 Astra.

Epoch says a model that can reason through complex problems without writing out its chain of thought may be harder to monitor, because safety teams catch problems by reading that chain. Epoch's method keeps adding steps a question requires until the model's accuracy is only 25 percentage points above chance, with each step weighted by difficulty (Epoch AI, 2026-11-17). This is a single measurement by Epoch, and no second team has repeated it. How we checked →

Source and checks

This item ran in the brief of Nov 18, 2026, alongside 14 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.

Cite this item

SecondSource (2026). "Epoch AI, an independent research group that runs its own benchmarks on AI models, found that GPT-6.2 Astra can chain 12.3 steps internally without writing out its reasoning, against 10.5 steps for the previous GPT-6 Astra.." https://secondsource.io/en/research/2026-11-18-1/. Accessed access date.