What happened

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November. Artificial Analysis tested Astra on questions that require it to find and combine information in financial reports and spreadsheets.

Why it matters

The benchmark gives credit only if the model answers the same question correctly on every attempt. This measures repeat accuracy on the same work.

The data

Artificial Analysis calls this measure pass^5. Each question has five independent attempts. The table shows three settings for GPT-6.2 Astra.

ItemValueUnitDate
Max62.5percent2026-11-19
Xhigh62.5percent2026-11-19
High61.3percent2026-11-19