Research
Researchers at Samaya AI released FrontierFinance, a fully open benchmark of 220 expert-crafted queries and 11,543 source-attributed rubrics spanning the full investment analyst workflow, built specifically because existing finance benchmarks have been largely saturated by current models
Where prior benchmarks mainly tested narrow financial data extraction, a task current frontier models have already mastered, FrontierFinance spans six use cases including financial modeling, sector and macro analysis, earnings and events, company research, catalyst monitoring, and screening, scored against thousands of source-attributed rubric items rather than generic LLM-as-judge grading. The authors position it as both broader and harder than existing public finance benchmarks, aimed at open-ended, long-form analyst queries that reference-based metrics struggle to score.
Benchmark saturation is a quiet but important story in AI-for-finance right now, and this paper is explicit about the problem it is responding to: the previous generation of finance benchmarks tested things like extracting a number from a 10-K, and current models have gotten good enough at that narrow task that the benchmarks stopped discriminating between good and mediocre systems. When every frontier model scores near-ceiling on a benchmark, the benchmark has stopped being useful for buyers trying to tell products apart, which is exactly the gap FrontierFinance is trying to fill with harder, more open-ended tasks.
The choice to grade against 11,543 source-attributed rubric items rather than a generic LLM-as-judge score is the more interesting methodological decision here, because it addresses a specific failure mode that has dogged this field: an LLM judging another LLM's financial analysis tends to reward confident, well-structured prose over factually grounded, source-traceable reasoning. Source attribution forces the evaluation to check whether a claim in the answer actually traces back to a specific piece of evidence, which is a much closer proxy for what a real research director wants from an analyst than fluency is.
Spanning the full investor workflow, financial modeling through screening and discovery, rather than a single task type, matters because it is closer to how AI research tools actually get evaluated by the firms buying them. Nobody is deploying an agent that only extracts numbers from filings; they are deploying something that has to reason across earnings calls, sector context and screening logic in the same session, and a benchmark that tests those capabilities in isolation was always going to overstate real-world readiness relative to one that tests them together.
The open release matters as much as the content. A fully open benchmark with public queries and rubrics means every finance-AI vendor claiming strong performance can now be checked against a shared, harder yardstick rather than marketing their own internal eval numbers, which has been a persistent transparency problem in this space. Whether FrontierFinance becomes the field's actual reference point depends on adoption, benchmarks only matter if labs and vendors report against them, but a benchmark built specifically because the old ones stopped being discriminating is a useful signal that the underlying models have genuinely moved past the last evaluation bar.
Read the original: arXiv - Researchers at Samaya AI released FrontierFinance, a fully open benchmark of 220 expert-crafted queries and 11,543 source-attributed rubrics spanning the full investment analyst workflow, built specifically because existing finance benchmarks have been largely saturated by current models. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.