Coding Agent Index Shows Tenfold Price Gap for Comparable Benchmark Scores

Benchmark Evaluation and Score Distribution The Artificial Analysis Coding Agent Index evaluates autonomous coding agents across three distinct coding benchmarks, assessing how model capabilities translate when paired with specific agent harnesses. In the latest benchmark sweep, Claude Sonnet 5.5 running in Claude Code captured the top position with an overall score of 68. Gemini 4 Argon running in Antigravity CLI followed with a score of 64, while GPT-6.1 Sol configured with extra high effort in Codex achieved a score of 63. ### Severe Cost Discrepancy Across Agent Harnesses Although benchmark performance clustered within a narrow five-point spread across all three systems, execution costs diverged significantly. Claude Sonnet 5.5 in Claude Code registered the highest measured expense at $14.19 per completed task. Gemini 4 Argon in Antigravity CLI averaged $5.84 per task, representing less than half the cost of Sonnet 5.5 (utilizing Google promotional rates). GPT-6.1 Sol in Codex recorded $1.04 per task, operating at roughly one-sixth the cost of Argon and one-fourteenth the cost of Sonnet 5.5. ### Practical Routing Recommendations for Teams For daily development and continuous agent execution, running Claude Code with maximum reasoning on repetitive chores creates compounding token invoices. Tech leads can implement multi-tier harness policies: dispatch high-volume bug fixing, refactoring, and test writing to GPT-6.1 Sol in Codex at $1.04 per task, and escalate failed trajectories or multi-file architectural overhauls to Claude Sonnet 5.5.