Our in-country experts set the standard. AI21 trained NVIDIAโs Nemotron on it.
The results: 8 points above Claude Opus 5.5 on our test set, in three weeks.
Read the full story: https://t.co/FQ5Yw8rj2d
Claude Opus 5.5 fully answered only 26% of our payroll questions.
With @AI21Labs and @nvidia, we built an open payroll model for @deel, the global payroll and HR platform.
It beats Opus, at 50x lower cost.
Hereโs how we did it ๐
AI learned from the open web.
It never learned how a payslip is calculated in Brazil, or how a contract is reviewed in Germany.
Emet, @deel's AI lab, brings in the experts who do this work every day. They turn it into benchmarks + training data for every model builder.
EmBench Global Work scores every answer fact by fact.
See the full leaderboard โ https://t.co/3s46s7J9Oa
Why it reflects real-world global work โ https://t.co/ik1BPrTJBB
Notice periods. Leave rules. Who can legally start work. We asked leading AI models 601 questions like these, with search tools. The best one got 75% of the facts right.
EmBench Payroll was built with 29 in-country payroll specialists. See the full leaderboard โ https://t.co/Jdv9EzvFfU
Why it reflects real-world payroll โ https://t.co/exjg80rwAE
The misses are small. A contribution cap, a filing deadline, a rule that changed this year. Close enough to look right. Wrong enough to break a paycheck.
Introducing Emet, the AI lab for model evaluation and training, specializing in how businesses run.
AI has read the whole internet. Itโs never run payroll or a business. Emet is built by Deel, which runs payroll, HR and compliance in 150+ countries.
First proof point: Pangaea, Deelโs family of LLMs, built on open-weight models and trained with Emet.
It's scored like every other model: on EmBench Global Work, it answers from memory within 5 points of the top model with search, up to 3x faster
EmBench Payroll: 295 real payroll tasks across 8 countries. The best model fully solves only 49%.
EmBench Global Work: 601 questions on working across borders. The best model, with search, gets 75% of the facts right.