GPT-6 Astra scores 52.5/100 on ILE forecast component

Investor's Last Exam, an equity-research benchmark described in an arXiv preprint published Oct. 30, ranked GPT-6 Astra at 68.1 out of 100 across 86 investment assignments.

The result matters for firms testing financial AI agents because Astra’s overall lead narrowed sharply in Forecast & Model, where it scored 52.5 out of 100 versus 68.1 overall, according to the preprint.

Claude Opus 5.5 placed second overall at 66.9 out of 100, the authors said. The paper lists Huangfei Jiang as lead author, with Anning Hu, Kai Sheng, Yuchen Cai, Rui Ge, Xiangyu Kong, Jianghao Lin, Shuzi Chen, Jiyi Weng and Junchi Yan as co-authors.

The benchmark is led by experienced buy-side investors and fund managers, according to the abstract. It is built around dated information, so agents are supposed to use only evidence available at the decision point and are then scored against later outcomes.

The authors wrote that ILE “shifts the evaluation unit from isolated capabilities to the evidence--assumption--judgment chain.” They said the design is intended to reveal where research skill fails to become usable forecasts or investment judgments.

That framing targets a common weakness in financial-AI testing. The preprint says existing benchmarks assess knowledge, research processes or forecasting separately, while full equity research needs a frozen record of what could have been known before the outcome.

ILE’s database contains 3.43 million dated document versions from nine source categories, the authors said. They also included structured market data and used verified time labels for both data lookup and calculations, limiting use of post-decision information.

The tasks are organized around deliverables used by professional analysts. The scoring system links criteria to source records, calculations and assumptions through financial dependency graphs, a structure that maps how evidence feeds an analytical conclusion.

The preprint is not yet peer reviewed. The arXiv abstract does not list the nine source categories, and the reported investment assessment is simulated rather than evidence of live portfolio performance.

The result also does not show that GPT-6 Astra or Claude Opus 5.5 cannot improve with different tools, prompts or training. It shows that, under ILE’s framework, higher overall research scores did not translate into the top forecast score disclosed in the abstract.

The paper was submitted to arXiv on Oct. 28 as version 1 under Computer Science > Artificial Intelligence. Its DOI is listed by arXiv as pending DataCite registration.

Source: arXiv preprint, arXiv cs.AI, Oct. 30, 2026