GPT-6 Astra Tops Investor's Last Exam, Stumbles on Forecasts

GPT-6 Astra led a new equity-research benchmark at 68.1 out of 100 overall but scored just 52.5 on its 'Forecast & Model' component, according to a paper posted to arXiv from a team including Huangfei Jiang and Junchi Yan. Claude Opus 5.5 trailed Astra narrowly on the overall mark, at 66.9.

The test, called Investor's Last Exam, is described in the abstract as 'led by experienced buy-side investors and fund managers' and spans 86 assignments drawn from a versioned financial database of 3.43 million document versions across nine source categories, with 'validated timestamps governing retrieval and computation.' The design is meant to stop agents from quietly borrowing hindsight from training data when they produce forecasts and investment judgments.

'Even state-of-the-art models fail to translate overall research capability into predictive performance,' the authors write. The paper argues the exam shifts the evaluation unit from 'isolated capabilities' to the full chain of evidence, assumption and judgment that is supposed to turn analyst work into a view that actually makes or loses money.

The abstract does not list scores for any models beyond Astra and Opus 5.5, and does not break the 52.5 Forecast & Model result into finer task categories.