Hermes Model Leaderboard Launches, Opus 5.5 Takes Top Spot, Astra Ranks Second But Is Priciest

Beating AI Express: Nous Research launches the Hermes Index, a model leaderboard for Hermes Agents, to compare the performance of different models within the Hermes ecosystem. All models run on the same Hermes Agent framework, with inference settings uniformly set to "high". The leaderboard calculates average scores across four benchmarks—Hermes Bench, Terminal-Bench 4, Terminal-Bench Science, and SkillsBench—alongside average cost per task. Hermes Bench is Nous Research’s new custom test, featuring 150 tasks spanning 25 categories covering skills, research, charts and creation, memory, tool use, and security. Agents process workspaces and real files directly, with scoring based on generated outputs, task status, and tool usage logs. The initial version tests 14 models. Claude Opus 5.5 tops the leaderboard with 63.31 points, followed by GPT 6 Astra at 56.25 points, and Claude Sonnet 5.5 at 53.14 points. Astra has the highest average cost per task at $11.61, while Opus 5.5 costs $4.99 per task and Sonnet 5.5 costs $2.82; some data for the latter two are still marked as provisional by their respective developers. Among low-cost models, DeepSeek V4.1 Flash scores 36.91 points at $0.259 per task, and GPT 6 Luna scores 33.89 points at $0.141 per task. Both models fall on the Pareto frontier of cost vs. performance, meaning no model exists that is cheaper while achieving a higher score.