# TaxEval v2

Hard tax questions, created with financial and tax experts. Scores answer correctness and stepwise reasoning against expert reasoning.

## Facts

- Maker: Vals AI (United States)
- Year: Not stated
- Use case: Finance knowledge
- Size: Over 1,500 questions (20 public validation, 300 private validation, 1,223 test)
- Public split: No public download; validation sets are available on request or for purchase
- Grading: Accuracy, judged by an LLM (Claude Sonnet 4.5, non-thinking, since 17 Nov 2025; Claude 3.5 Sonnet before).
- Licence: proprietary
- Submission: Not stated
- Products allowed: Not stated
- Latest version: v2
- What it does not cover: Archived as saturated: no longer run on new models. No tool use, retrieval or citations (see Tax Agent Bench for agentic tax research).
- Last checked: 2 Nov 2026
- Links: site https://www.vals.ai/benchmarks/tax_eval_v2, leaderboard https://www.vals.ai/benchmarks/tax_eval_v2
- Directory page: https://benchmarks.finance/benchmarks/vals-tax-eval-v2/
- Leaderboard: https://benchmarks.finance/leaderboards/vals-tax-eval-v2/

## Results

### TaxEval v2: accuracy % (v2) (higher is better, 5 models)

| Rank | Model | Provider | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Muse Spark 1.2 | Meta | 80.4 | accuracy % (v2) | test set (1,223 questions), archived board | Official board | Vals.ai TaxEval v2 leaderboard | https://www.vals.ai/benchmarks/tax_eval_v2 | 1 Sept 2026 (checked 2 Nov 2026) |
| 2 | Muse Spark 1.1 | Meta | 79.7 | accuracy % (v2) | test set (1,223 questions), archived board | Official board | Vals.ai TaxEval v2 leaderboard | https://www.vals.ai/benchmarks/tax_eval_v2 | 1 Sept 2026 (checked 2 Nov 2026) |
| 3 | Muse Spark | Meta | 77.7 | accuracy % (v2) | test set (1,223 questions), archived board | Official board | Vals.ai TaxEval v2 leaderboard | https://www.vals.ai/benchmarks/tax_eval_v2 | 1 Sept 2026 (checked 2 Nov 2026) |
| 4 | Claude Sonnet 4.6 | Anthropic | 77.1 | accuracy % (v2) | test set (1,223 questions), archived board | Official board | Vals.ai TaxEval v2 leaderboard | https://www.vals.ai/benchmarks/tax_eval_v2 | 1 Sept 2026 (checked 2 Nov 2026) |
| 5 | Claude Fable 5 | Anthropic | 76.9 | accuracy % (v2) | test set (1,223 questions), archived board | Official board | Vals.ai TaxEval v2 leaderboard | https://www.vals.ai/benchmarks/tax_eval_v2 | 1 Sept 2026 (checked 2 Nov 2026) |


Source: https://benchmarks.finance/benchmarks/vals-tax-eval-v2/

Created and maintained by Moneda.