<!-- LLM_VERSION_INFO
FORMAT: text/markdown
CONTENT_TYPE: article
ORIGINAL_URL: https://www.vals.ai/benchmarks/tax_eval_v2
ALTERNATE_VERSION: benchmarks/tax_eval_v2/index.html (text/html)
EXTRACTION_DATE: 2026-04-17T00:42:48.942Z

This is the markdown version with text-only content (images converted to alt-text).
For rich formatting with images, request the HTML version at: benchmarks/tax_eval_v2/index.html
-->

# TaxEval (v2)

Proprietary

Vals AI original benchmark crafted with industry experts on non-public datasets

Updated: 4/16/2026

A Vals-created set of questions and responses to tax questions

## Models Overview

| Model                                | Accuracy        | Cost In/Out   | Latency  |
|--------------------------------------|-----------------|----------------|----------|
| [Muse Spark](/content/models/meta_muse_spark/index.html)          | 77.68%± 0.81    | N/A            | 57.52s   |
| [Claude Sonnet 4.6](/content/models/anthropic_claude-sonnet-4-6/index.html)  | 77.11%± 0.82    | $3 / $15      | 127.88s  |
| [Claude Opus 4.6 (Thinking)](/content/models/anthropic_claude-opus-4-6-thinking/index.html) | 75.96%± 0.83    | $5 / $25      | 79.25s   |
| [Grok 3](/content/models/grok_grok-3/index.html)                  | 75.88%± 0.83    | $3 / $15      | 13.06s   |
| [GPT 5.2](/content/models/openai_gpt-5.2-2025-12-11/index.html)         | 75.76%± 0.85    | $1.75 / $14   | 60.48s   |
| [Grok 4 Fast (Reasoning)](/content/models/grok_grok-4-fast-reasoning/index.html) | 75.70%± 0.76    | $0.2 / $0.5   | 10.18s   |
| [Claude Opus 4.7](/content/models/anthropic_claude-opus-4-7/index.html)   | 75.27%± 0.84    | $5 / $25      | 62.07s   |
| [GPT 5 Mini](/content/models/openai_gpt-5-mini-2025-08-07/index.html)     | 75.22%± 0.85    | $0.25 / $2    | 37.43s   |
| [GPT 4.1](/content/models/openai_gpt-4.1-2025-04-14/index.html)           | 75.06%± 0.84    | $2 / $8       | 7.89s    |
| [GPT 5.1](/content/models/openai_gpt-5.1-2025-11-13/index.html)           | 74.86%± 0.85    | $1.25 / $10   | 44.18s   |

See 94 more models

## Key Takeaways

- [Muse Spark](/content/models/meta_muse_spark/index.html) leads with 77.68% accuracy, narrowly edging out [Claude Sonnet 4.6](/content/models/anthropic_claude-sonnet-4-6/index.html) at 77.11% and [Claude Opus 4.6 (Thinking)](/content/models/anthropic_claude-opus-4-6-thinking/index.html) at 75.96%.
- The top models are all tightly clustered, with the top ten separated by less than three percent.

## Dataset and Context

TaxEval v2 evaluates models’ abilities to answer hard tax-related questions. This version focuses on both answer correctness and structured reasoning capabilities.
This dataset was created in collaboration with financial and tax experts, who have both created and double-checked all questions and answers.

Some key features:

- 1,500+ total questions across validation and test sets
- A balanced distribution of topics and question types
- Comprehensive evaluation of both answers and reasoning steps

The benchmark consists of three main components:

1. **Public Validation Set**: 20 samples, available upon request (contact [contact@vals.ai](mailto:contact@vals.ai)).
2. **Private Validation Set**: 300 samples for model evaluation, available for purchase to evaluate and improve models.
3. **Test Set**: 1,223 samples. These samples are never shared.

The benchmark is composed of two tasks (each task uses the same questions, but they are evaluated differently):

1. **Answer Correctness**: The factual correctness of the answer, as compared to a ground truth.
2. **Stepwise Reasoning**: The quality and structure of the reasoning process as compared to the reasoning process of human experts. This ensures models not only provide correct answers but also demonstrate a clear thinking process.

The benchmark includes a diverse range of question types:

- Application and Compliance (18.3%)
- Comparative Analysis (16.2%)
- Numerical Reasoning (16.7%)
- Problem Solving and Critical Thinking (16.5%)
- Semantic Analysis (18.0%)
- Updates and Current Affairs (15.9%)

Each category is carefully balanced between the private validation and test sets to ensure representative sampling. The evaluation process uses [Claude Sonnet 4.5 (Nonthinking)](/content/models/anthropic_claude-sonnet-4-5-20250929/index.html) as judge for both answer correctness and stepwise reasoning assessment.

## Results

### Model Output Examples

The hardest questions for the models typically require complex, multi-step calculations, reasoning on what laws and numbers to use given the information, or accessing more recent data.

Below is an example of a question where one model arrives at the correct answer ($18,280) while the others perform the calculations incorrectly, despite all models using a significant amount of reasoning tokens.

#### Michael, a married taxpayer filing jointly, has an adjusted gross income (AGI) of $300,000 for 2023, excluding any investment income. During the year, he received $20,000 in interest from corporate bonds and $15,000 in interest from municipal bonds issued by his state of residence. In addition, he sold a collectible artwork for $100,000 that he had purchased 5 years ago for $60,000. Calculate Michael's tax liability related to his investment income, including any applicable taxes on capital gains and considering the Net Investment Income Tax (NIIT).

**Ground Truth**  
**GPT 5**  
**Claude Sonnet 4.5 (Thinking)**  
**Grok 4 Fast (Reasoning)**

**Response:**  
Michael owes a total of $18,280 in taxes related to his investment income.

## Changelog

> This benchmark was updated on 11/17/25 to switch the LLM-as-judge from Claude 3.5 Sonnet, which had been deprecated and is no longer usable, to Claude 4.5 Sonnet.