Canadian Tax LLM Bench
Every model answers the same 130 questions, 10 for each province and territory, on figures a preparer looks up: provincial credits, transfer and fuel taxes, payroll taxes, benefits and corporate tax. Run 2026-10-15.
Leaderboard
| # | Model | Lab | Score | Cost of one run |
|---|---|---|---|---|
| 1 | GPT-6 Astra | OpenAI | 80.8 ±14.7 | $1.08 |
| 2 | Claude Opus 5.5 | Anthropic | 53.8 ±17.9 | $0.53 |
| 3 | Claude Fable 5.1 | Anthropic | 53.8 ±17.9 | $1.74 |
| 4 | GPT-6 Sol | OpenAI | 30.8 ±16.7 | $0.29 |
| 5 | DeepSeek V4.1 Flash | DeepSeek | 23.1 ±15.5 | $0.20 |
| 6 | Gemini 3.8 Flash | 19.2 ±14.7 | $0.063 | |
| 7 | GPT-6 Luna | OpenAI | 19.2 ±14.7 | $0.019 |
| 8 | Gemini 3.1 Pro | 15.4 ±13.7 | $0.52 | |
| 9 | Claude Sonnet 5 | Anthropic | 15.4 ±13.7 | $0.52 |
| 10 | GLM-5.3 Flash | Z.ai | 7.7 ±11.0 | $0.025 |
| 11 | GLM-5.3 | Z.ai | 3.8 ±9.1 | $0.32 |
Sample questions
Real questions from the bench, with the right answer and what every model replied: all 4 questions and answers.
How the bench works
Every answer key is computed by code from a table of official 2025 figures (Canada Revenue Agency, Revenu Québec, provincial finance ministries, workers' compensation boards), and each table was built twice by agents that never saw each other's work. A dollar answer must match to the cent. Quebec questions are asked in French. This is a test of models, not tax advice.