Research

An independent benchmark scores 59 models on realistic financial-analyst work, and nobody clears 62%

Finance Agent v2, a severity-weighted, repeated-run benchmark of qualitative analysis, quantitative work, comparables, precedents, earnings and modeling tasks, puts Google's Gemini 3.8 Flash on top at 61.4% as of 10 September 2026, with Claude Fable 5.1 and Claude Opus 5 close behind in the high 50s.

· Source: BenchLM.ai


Most of the capability claims this site tracks arrive as vendor announcements or single-firm case studies, Bracket22 running a book on three named agents, Morgan Stanley opening a wealth-management funnel to external AI, JPMorgan promising longer-running agents later this year. Finance Agent v2 is a different kind of data point because it isn't anyone's launch story: it's a standing, third-party leaderboard that runs 59 models against the same set of realistic analyst tasks, qualitative analysis, quantitative work, comparables, precedents, earnings, disclosure, modeling, and scores them with severity-weighted partial credit across repeated runs rather than a single best-of pass. The top score as of this week's snapshot is 61.4%, held by Gemini 3.8 Flash. The next four models, including Claude Fable 5.1 and Claude Opus 5, sit within about three points of each other in the high 50s.

The number worth sitting with is the ceiling, not the ranking. This is a professional-grade benchmark built to look like the work an analyst actually does, not a synthetic reasoning puzzle, and the best model available still gets roughly two out of every five graded task-elements wrong or partially wrong. That is a meaningfully different picture from the fluency these systems display in a demo or a well-scoped single query, and it's the gap that keeps showing up whenever this site has looked past a headline capability claim: Bracket22's Brian Kelly frames his three agents as specialists whose synthesis he personally reviews before every trade, not as a system he trusts to work unsupervised, and the Goldman Sachs partner quoted alongside that story worried specifically about losing the human pattern-recognition needed to catch a plausible-looking wrong answer. A 61% top score on realistic analyst tasks is the quantitative version of exactly that worry.

It also complicates the framing that agentic capability is a smoothly rising line firms can plan a deployment calendar around. Finance Agent v2's rankings cluster models from different labs, different architectures and different release dates within a few points of a mid-50s to low-60s band, which reads less like a capability that's about to break through a ceiling and more like a genuine current limit of what these systems do well on financial-analyst-shaped work specifically, as opposed to coding or general reasoning benchmarks where frontier models have posted much higher scores. For a reader deciding how much weight to put behind an agent's output on comparables or earnings analysis, the honest read of this snapshot is that the best available model is still wrong on a large minority of the work often enough that the review step, exactly the step Kelly and Morgan Stanley's own client-facing design both keep in place, is not a formality. It is currently carrying a meaningful share of the accuracy the model itself doesn't provide.


Read the original: BenchLM.ai - An independent benchmark scores 59 models on realistic financial-analyst work, and nobody clears 62%. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.