A haunted house but it's just a manual hiring process
A haunted house but it’s a spreadsheet called “Q3 expenses (real)”
A haunted house but it’s a perfect candidate and “we don’t hire there”
GPT-6 Astra is the top model for payroll according to PayMark, our new AI payroll benchmark.
It got 26.0% of 2,000 payslips fully right, ahead of Claude Opus 5.5 at 25.9%. For reference, the top four models are within 2.5 points of each other.
Accuracy means when net pay,
Claude Fable 5.1 tops HRmark, our employment-law benchmark, at 69.0%.
Gemini 3.8 Flash scored 67.2% at $0.76 per 1k answers, vs $49.74 for Fable.
High effort moved the median 0.6 pts for 2.7x the cost.
Check the accuracy before you pay for the extra reasoning. Full results ↓
We built two benchmarks to test AI on global HR and payroll: HRmark (employment law) and PayMark (payroll).
12 models, 8 providers, 50 countries.
Best on HRmark: Claude Fable 5.1, 69%.
Best on PayMark: GPT-6 Astra, 26% of payslips fully accurate.
bit.ly/4zcYUWd
The world works because people work.
Shot in Perugia. Edited in New York. Sourced in Berlin. Paid in Singapore.
Our new campaign, For a World That Works, celebrates the unseen infrastructure powering the global economy.
#PayTheWorld
