Six coding agents now lead the Artificial Analysis Coding Agent Index, a composite leaderboard built from three separate benchmarks. This index is designed for engineers comparing agents across software engineering tasks, measuring not just task success but also cost, token efficiency, and agent runtime. If you’re deciding which agent to try next, this is the ranking most teams are looking at.
Each agent’s headline score is an average across three distinct benchmarks and three attempts per task, but the underlying tests cover very different ground. The composite figure is useful, but understanding what it actually measures—and what it omits—matters if you want to avoid surprises in production.
What the benchmarks actually test
The Coding Agent Index combines three public benchmarks, each focused on a different aspect of agent capability:
FrontierCode-AA 2.0covers 132 tasks involving software engineering implementation and bug-fixing, developed by Cognition and implemented by Artificial Analysis.Terminal-Bench 4.1includes 66 tasks that stress terminal workflows—shell navigation, tool execution, and working through multi-step command-line scenarios. This comes from the Laude Institute.SWE-Atlas-QnA-AAfeatures 124 tasks in technical Q&A and repository understanding, requiring an agent to read codebases, grasp architecture, and provide correct technical answers. This benchmark is by Scale AI, with grading adjustments by Artificial Analysis.
Each of these benchmarks is scored independently, and the index averages the results for a single number. The source notes that agents with similar composite scores can still differ significantly depending on which type of task you care about.




