FinBalance - Replayed Balance Sheet: leaderboard
FinBalance (2026) is a multi-document accounting reconciliation benchmark: from a bundle of source documents with distractors, a model posts journal entries citing their documents, aggregates them into a balance sheet and flags inconsistencies, scored against a deterministic ledger. This board replays the journal entries the model posted through the reference ledger and reports, in percent of the 480 standard records, how often the resulting balance sheet matches the ground truth; higher is better.
Metric: Replayed balance-sheet accuracy (%): share of records whose balance sheet, obtained by replaying the model's own journal entries through the reference ledger, matches the ground truth (BS recon); 480 standard records of the 710-record core evaluation split (source-document bundles with distractors across eight industries, three period types and five difficulty levels), baseline prompt, temperature 0, one pass; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Minimal) | 54.8 | #164 (Gemini 3 Flash) |
| 2 | GPT-5 (Low) | 46.3 | #167 (GPT-5) |
| 3 | Grok 4.3 (Non-reasoning) | 35 | #333 (Grok 4.3) |
| 4 | Qwen 3 235B A22B 2507 Instruct | 30.2 | #619 |
| 5 | Claude Haiku 4.5 | 29.2 | #568 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=finbalance-replayed-balance-sheet · How It Works · Data refreshed daily, snapshot 2026-11-13.