Nous Research launches Hermes Index to compare agent performance and cost

Nous Research has introduced Hermes Index, a leaderboard measuring how AI models perform within its Hermes Agent software. The company said the project aims to help users select models and give developers feedback on their performance.

The index averages results from Hermes Bench, TerminalBench 4, TerminalBench Science and SkillsBench. It also displays average spending per task. Hermes Bench contains 150 tasks covering 25 categories, with evaluations examining completed work and changes to workspace files.

Testing uses the same agent framework for each model, with high reasoning effort enabled when available. This provides a comparison of performance within Hermes, alongside the expense of running those tasks. The published methodology excludes four GPU tasks from TerminalBench 4 and removes affected SkillsBench tasks when models accessed publicly available solutions.

Some results remain provisional because testing is incomplete or costs are estimated. Those entries await full reruns, while the comparison interface already lets users examine individual benchmark scores and costs for up to four models.