AA Index · data 23.09.2026vibecoding.tech_

AI agent ranking

AA Coding Agent Index · artificialanalysis.ai · captured 23.09.2026 · 6 pairs

AI agents Models MCP

Agent + model

Claude CodeFable 5.1 · max*

62.2

$12.39

34.8 min

CodexGPT-6 Astra · max

61.6

$7.47

29.4 min

Claude CodeOpus 5 · max

59.7

$10.79

41.9 min

CodexGPT-6 Sol · max

56.7

$2.99

22.3 min

Grok BuildGrok 4.7 · xhigh

56.3

$8.82

39.2 min

OpenCodeGLM-5.3

53.6

$4.24

48.1 min

Claude CodeOpus 4.8 · max

—

CodexGPT-5.5 · xhigh

—

Claude CodeSonnet 5 · xhigh

—

Claude CodeHaiku 4.5 · xhigh

—

Claude CodeOpus 5 · xhigh

—

Claude CodeFable 5 · max*

—

Claude CodeGLM-5.3

—

CodexGPT-5.6 Sol · max

—

CodexGPT-5.6 Terra · max

—

Grok BuildGrok 4.6 · high

—

CodexGPT-5.6 Luna · max

—

Claude CodeSonnet 5 · max

—

Grok BuildGrok 4.5 · high

—

Kimi Code CLIKimi K3

—

OpenCodeGemini 3.8 Flash · high

—

Grok BuildGrok 4.6 · xhigh

—

Claude CodeDeepSeek V4 Pro · high

—

Claude CodeGLM-5.2

—

Claude CodeOpus 4.8 · medium

—

CodexGPT-5.5 · medium

—

020406080100

aa index + task price/time - artificialanalysis.ai (23.09) · t-bench 4.0 - tbench.ai (23.09) · deepswe v1.1 (22.09) · rubench r1 - /rubench (09.07) · "—" - pairing was not run · * - run with fallback model substitution, analysis - /rubench

How to read the table

the pair and the model are on different tabs

AI agents are measured as an agent+model pair (config in the badge), models on their own; the best config for each, one row.

the benchmark is selected on the right

the bar and rating column follow the selected benchmark; «—» - not run on it, the row moves down. Sort by clicking a column.

price and time - real

averages per task: agents - artificialanalysis measurement, models - the actual deepswe run bill. Not derived from tokens.

Index methodologywhere the numbers come from and why there is one row per pairRuBenchour benchmark on Russian technical specifications: pass@1 with open trajectoriesAll rankingsmodels, coding agents, MCP