# Artificial Analysis Expands Coding Agent Index to 322 Tasks in V2.1 By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/57087 Published: 2026-11-14T00:23:26.000Z Updated: 2026-11-14T00:23:26.000Z Section: Technology > The update replaces a 100-task FrontierCode set with a 132-task version and updates Terminal-Bench verifiers while keeping its 66 tasks. Artificial Analysis has expanded its Coding Agent Index to 322 tasks in version 2.1, a 32-task increase that reflects a revised benchmark mix and limits direct comparisons with earlier scores. The current index includes 132 FrontierCode-AA 2.0 tasks, 66 Terminal-Bench 4.1 tasks and 124 SWE-Atlas-QnA-AA tasks. The previous edition combined 100 FrontierCode v1.1 tasks with the same 66 Terminal-Bench and 124 SWE-Atlas-QnA-AA tasks, for 290 total. FrontierCode-AA 2.0 evaluates agents on changes to real open-source repositories, using maintainer-oriented criteria to judge whether the work is suitable for merging. Terminal-Bench covers tasks in command-line environments. Its 4.1 update keeps the same 66 tasks but revises the checks used to verify results. SWE-Atlas-QnA-AA tests whether an agent can trace code and explain how a repository behaves. Artificial Analysis runs each task three times, averages the attempts for that task, then averages task scores within each benchmark. The three benchmark scores count equally toward the index, which combines repository changes, terminal work and codebase questions. The results measure agent variants, so scores can reflect the model as well as the software harness and settings used to run it. The index’s overall score summarizes performance across different kinds of work rather than coding output alone. Mistral Large 4’s 48% result on the [Coding Agent Index]() came from an earlier edition. That result came from an earlier edition, and the revised benchmark mix means it is not a clean before-and-after comparison with a version 2.1 score. Version 2.1 replaces the earlier FrontierCode component and updates Terminal-Bench verification. Since the benchmark suite can change, score differences between editions do not by themselves show whether an agent improved or declined. ## Links in this article - [Coding Agent Index](https://www.tokenpost.com/news/technology/43784)