How the index works

Artificial Analysis runs each coding agent with a specific model and reasoning setting, so the same agent can appear several times. The Coding Agent Index v1.5 is the equal-weight average of three benchmarks: DeepSWE v1.1, SWE-Atlas-QnA and Terminal-Bench 4.0. Cost per task is calculated from per-token API prices, including cache pricing. It is not a subscription price, and many people use these agents through subscription plans instead. The figures here are from the Artificial Analysis leaderboard as checked on October 21, 2026 (Beijing time). Tiers are grouped by clear gaps between scores. Only scores within about one point of each other are described as tied, which applies to Tier 1.

Tier 1: tied at the top (68.0 to 68.4)

AgentModel and settingIndexCost per taskAvg. time per task
Claude CodeClaude Sonnet 5.5 (max)68.4$14.1987 min
Devin Fusion CLIClaude Opus 5.5 (max) + GPT-6 Luna (high)68.2$9.0255 min
Muse CodeMuse Wonder Full (max)68.0not published29 min

Devin Fusion pairs a lead model with a cheaper helper model, and it is the least expensive of the top three that have a published cost. Our brief on Devin Fusion's Coding Agent Index result covers that setup in more detail.

Tier 2: 66.0

AgentModel and settingIndexCost per taskAvg. time per task
Claude CodeClaude Opus 5.5 (max)66.0$13.0464 min

Tier 3: 61.6 to 63.8

AgentModel and settingIndexCost per taskAvg. time per task
Antigravity CLIGemini 4 Argon (high)63.8$5.8435 min
Muse CodeMuse Wonder (max)63.7$8.0422 min
Claude CodeClaude Fable 5.1 (max), with fallback62.2$12.3935 min
CodexGPT-6 Astra (max)61.6$7.4729 min

Tier 4: around 60

AgentModel and settingIndexCost per taskAvg. time per task
CodexGPT-6.1 Sol (max)60.1$1.5524 min

Tier 5: 51.9 to 56.3

AgentModel and settingIndexCost per taskAvg. time per task
Grok BuildGrok 4.7 (xhigh)56.3$8.8239 min
OpenCodeGLM-5.3 (max reasoning)53.6$4.2448 min
Kimi Code CLIKimi K3 (setting not listed)51.9$5.0561 min

Tier 6: below 45

AgentModel and settingIndexCost per taskAvg. time per task
Muse CodeMuse Spark Open (xhigh)41.2$3.3325 min
Claude CodeClaude Haiku 5.5 (max)36.8$2.8088 min
MirrorBeam 1.027.7not published42 min

Lower settings can be much cheaper

The tiers above use each agent's default entry. Artificial Analysis also tests other settings, and some score at Tier 3 level at much lower cost:

AgentModel and settingIndexCost per task
Claude CodeClaude Sonnet 5.5 (xhigh)62.9$3.33
CodexGPT-6.1 Sol (xhigh)62.9$1.04
CodexGPT-6.1 Sol (medium)61.4$0.70

For a closer look at Claude Code and Codex, see our Claude Code vs Codex comparison. For the difference between OpenAI's two GPT-6 models, see our comparison of GPT-6 Sol and GPT-6 Luna.

Sub-benchmark scores for the top tiers

Agent and modelDeepSWE v1.1SWE-Atlas-QnATerminal-Bench 4.0
Claude Code, Claude Sonnet 5.5 (max)72.0%66.9%66.2%
Devin Fusion CLI, Claude Opus 5.5 (max) + GPT-6 Luna (high)72.6%69.4%62.6%
Muse Code, Muse Wonder Full (max)77.0%68.0%59.1%
Claude Code, Claude Opus 5.5 (max)68.4%66.4%63.1%
Antigravity CLI, Gemini 4 Argon (high)78.8%56.5%56.1%
Muse Code, Muse Wonder (max)75.8%67.2%48.0%

Antigravity CLI with Gemini 4 Argon (high) has the highest DeepSWE v1.1 score among default entries, but lower SWE-Atlas-QnA and Terminal-Bench 4.0 scores keep it in Tier 3.

We will update this page when Artificial Analysis updates the leaderboard.