TaxCalcBench v2: Open-Source Benchmark Reveals How Frontier AI Models Handle Complex Tax Filing

Summary

Researchers have released TaxCalcBench v2, an open-source evaluation framework that tests frontier AI models on their ability to accurately file federal and state tax returns. The benchmark, updated in June 2026 for Tax Year 2025, includes realistic tax form inputs (W-2s, 1099s) and increasingly complex financial scenarios that challenge current large language models.

The v2 benchmark evaluated eight leading AI models with varying performance levels. OpenAI's GPT-5.5 achieved the highest accuracy at 54% for complete correct returns (strict metric) when equipped with web search capabilities, while Claude Fable 5 with web search reached 34% and Claude Opus 4.8 achieved 30%. Without web search access, performance dropped significantly across the board, with GPT-5.5 managing only 24% correct returns and Claude models scoring between 6-26%.

The benchmark also revealed that even frontier models struggle with complex tax scenarios: even the top performer only correctly filed roughly half of the test cases. Performance measured at the individual tax form line level showed better results, with GPT-5.5 achieving 88.89% accuracy (lenient) on line-by-line calculations, suggesting models can handle individual calculations but struggle with form selection, deduction logic, and cross-form dependencies.

The v2 release substantially expands complexity compared to prior versions, covering multi-state returns and more intricate financial situations. The results underscore that while AI models have made significant progress on technical capabilities, accurately filing taxes—requiring careful form selection, regulatory knowledge, and attention to edge cases—remains a frontier challenge for frontier AI systems.

• TaxCalcBench v2 significantly increased complexity over prior versions with realistic PDFs, state returns, and intricate multi-situation tax cases, making it a more realistic evaluation of AI capability for real-world tax applications

Editorial Opinion

TaxCalcBench v2 provides valuable empirical evidence that frontier AI models, despite remarkable capabilities in other domains, struggle meaningfully with tax filing—a task requiring regulatory knowledge, careful attention to rules, and systematic logic. The stark gap between line-level accuracy and complete-return accuracy is particularly revealing: models can calculate correctly but fail at the higher-level reasoning needed to route information to the right forms and apply conditional rules. This benchmark demonstrates that capability progress isn't uniform across domains; tax filing remains a meaningful test of AI reasoning and knowledge integration that current frontier models have not yet solved.