Inside Vals AI's CorpFin v2: why credit-agreement AI is still hard, how the Vals Index ranks the top models, and what it means for corp-dev and credit teams.
Vals AI's CorpFin v2 is the benchmark closest to our own world: it feeds models public commercial credit agreements and asks expert-written questions that require extracting terms and reasoning across the document. The question set, 858 test questions across 43 documents, drafted by financial analysts, legal professionals and academics, is designed around how leveraged-finance professionals actually read contracts.
That design exposes exactly where models struggle. Market-standard concepts such as "baskets" carry commonly understood meanings that are almost never explicitly labelled in the agreement itself; one sample question asks whether a contract contains a "Chewy blocker," a clause preventing a subsidiary's release from its debt obligations. The hardest questions require synthesizing provisions and calculating values across multiple sections, and PDF-parsed text, special characters, tables, defined-term cross-references, adds a further layer of failure modes.
For a composite view, Vals AI publishes the Vals Index, weighting finance benchmarks (CorpFin and Finance Agent) at 2.0 and coding at 1.4. The public snapshot has GPT-5.5 on top at 67.62%, Claude Opus 4.7 at 66.10% and Gemini 3.5 Flash at 62.05%, while SpaceXAI's newly released Grok 4.5 debuted at 65.30%, up nearly 19 points from Grok 4.3, the sharpest single-generation improvement on the index.
Our Take
for corp-dev and credit teams, independent evals like CorpFin v2 are the right diligence instrument, they convert vendor claims into task-level evidence on documents that look like your documents. Two lessons follow. First, leaderboard positions churn with every release cycle, so architectures should stay model-agnostic. Second, sub-70% composite scores mean covenant extraction and comparable analysis still require expert review; the models are assistants on credit work, not yet arbiters.