CorpFin v2 — Results
| Rank | Model | Score |
|---|---|---|
| 1 | claude-opus-5 | 73.2 |
| 2 | claude-opus-5 | 73.2 |
| 3 | claude-fable-5 | 71.8 |
| 4 | claude-fable-5 | 71.8 |
| 5 | kimi-k3 | 71.6 |
| 6 | muse-spark-1-2 | 70.9 |
| 7 | muse-spark-1-2 | 70.9 |
CorpFin v2
1 phaseActive
Vals AI private benchmark on understanding long-context credit agreements: 858 test questions from 43 documents. Archived by Vals AI after performance saturated. Metric: % accuracy.
Overview
CorpFin v2
Archived: Vals AI no longer runs CorpFin v2 on new model releases because performance on it has saturated. Its last leaderboard update was 2026-08-12, so newer models have no CorpFin v2 score.
Quick answer: CorpFin v2 is a Vals AI benchmark that tested how well AI models understand long credit agreements, often more than 200 pages. It has 858 private test questions from 43 documents. The final leaderboard is led by Claude Opus 5 at 73.2%, with the top models within about two points of each other.
At a Glance
What it tests: Reading and answering questions about long-context credit agreements, the dense contracts behind corporate loans.
Why it matters: Credit agreements are long, cross-referenced legal and financial documents; answering questions about them tests long-context retrieval and careful reading rather than general knowledge.
Known limitations: The benchmark is archived and saturated: the top models score within about two points of each other, so it no longer separates frontier models. The test set is private, so outside teams (Benchgen included) cannot reproduce the scores; they can only be mirrored from Vals AI.
What CorpFin v2 Measures
Each question is about a specific credit agreement and has to be answered from that document. The agreements often run past 200 pages, so models have to find the relevant clauses across a long context and reason over them accurately.
Benchmark Specifications
| Field | Value |
|---|---|
| Task category | Reasoning / long-context finance documents |
| Questions | 858 test questions from 43 documents (plus 340 private validation and 20 public validation questions) |
| Metric | % accuracy |
| Test set | Private: "the privately held set for Vals AI benchmarking and is never shared" |
| Public set | 20 questions from 1 document, available from Vals AI on request |
| Saturation | High |
| Status | Archived by Vals AI (last updated 2026-08-12) |
| Created by | Vals AI |
| Leaderboard | vals.ai/benchmarks/corp_fin_v2 |
How CorpFin v2 Is Scored
Models answer questions about each credit agreement, and the score is the share of test questions answered correctly.
State-of-the-Art Results
Final Vals AI leaderboard, last updated 2026-08-12:
| Rank | Model | Score | Source | Date |
|---|---|---|---|---|
| 1 | Claude Opus 5 | 73.2% | Vals AI | 2026-08 |
| 2 | Claude Fable 5 | 71.8% | Vals AI | 2026-08 |
| 3 | Kimi K3 | 71.6% | Vals AI | 2026-08 |
| 4 | Muse Spark 1.1 | 71.3% | Vals AI | 2026-08 |
| 5 | Muse Spark 1.2 | 70.9% | Vals AI | 2026-08 |
The Kimi K3 score on the Benchgen leaderboard (71.6%) matches its Vals AI result.
CorpFin v2 on Benchgen
No Benchgen runs. The test set is private and the benchmark is archived, so the scores above are mirrored from Vals AI and will not be updated.
CorpFin v2 vs Other Benchmarks
| Benchmark | What it tests | Saturation |
|---|---|---|
| CorpFin v2 | Long credit agreements (archived) | High |
| Finance Agent v2 | Analyst questions over public company filings | Low |
| Harvey Lab-AA | Legal knowledge-work quality | Low |
| SpreadsheetBench 2 | Spreadsheet manipulation & formula tasks | Low |
Run CorpFin v2 on Your Model
The official test set is private and Vals AI has stopped running the benchmark, so there is no way to get a new comparable score. For current finance results, use Finance Agent v2. On Benchgen you can build a long-document evaluation from your own contracts and track your model against it.