Which AI Coding Agent Reigns Supreme in 2026? A Comparative Ranking of Cursor, Claude Code, and Codex

Overview

Artificial Analysis released a horizontal evaluation of AI coding agents covering four dimensions: performance, token consumption, cost, and execution time.

Benchmarks

SWE‑Bench‑Pro‑Hard‑AA – real‑world bug‑fix scenarios.

Terminal‑Bench v2 – tool‑chain usage.

SWE‑Atlas‑QnA – code‑base understanding.

Agents and Scores

Agents tested include Claude Code, Cursor CLI, OpenAI Codex, Google Gemini CLI and model back‑ends such as Opus 4.7, Sonnet 4.6, GLM‑5.1, Kimi K2.6 and DeepSeek V4 Pro.

Comprehensive ranking : Cursor CLI + Opus 4.7 achieved 61 points. Codex (GPT‑5.5) and Claude Code (Opus 4.7) each scored 60 points, leaving only a one‑point gap among the top three.

Token Consumption

The highest token usage was observed for Claude Code + GLM‑5.1 at 4.8 M tokens per task, roughly three times the 1.5 M tokens consumed by Cursor CLI + Opus 4.7. Most tokens were spent on cache hits (shown in orange in the original charts), which significantly reduces the actual cost.

Cost per Task

The cheapest configuration was Cursor CLI + Composer 2 at $0.07 per task. The most expensive was Claude Code + GLM‑5.1 at $2.26 per task, a 32‑fold difference. DeepSeek V4 Pro cost $0.35 per task and earned 50 points, saving nearly three‑quarters of the cost of the Opus 4.7 setup while incurring only about a 17 % performance loss.

Execution Time

The fastest setup was Claude Code + Opus 4.7 (direct Anthropic connection) at 5.8 minutes per task, achieving a score of 60. In contrast, Claude Code + Kimi K2.6 required 41.5 minutes for the same 50 points, likely due to a longer inference chain and higher API latency.

Key Observations

The top three agents are separated by only one point, indicating diminishing returns from pure performance ranking.

Future differentiation will focus on cost efficiency and latency.

Domestic models (DeepSeek V4 Pro, Kimi K2.6, GLM‑5.1) demonstrate viable capabilities but exhibit higher token consumption and slower response times.

DeepSeek V4 Pro emerges as the most attractive cost‑performance option.

Reference

https://artificialanalysis.ai/agents/coding-agents

Code example

往
期
推
荐
1、
十万个Why：两个字段都建了单列索引，为什么加了 OR，执行计划还是全表扫描？
2、
Java 枚举别只会写 SUCCESS/FAIL！这 5 种高级 Enum 玩法，顶级框架早就在偷偷用了
3、
Java 枚举别只会写 SUCCESS/FAIL！这 5 种高级 Enum 玩法，顶级框架早就在偷偷用了
4、
后知后觉：腾讯TIM PC客户端竟然已经被弃用 显示版本过低无法登录
5、
IntelliJ IDEA 从卡顿到起飞，只用改这些。。。
6、
我说：“咱们以后面试，必须要问 Harness” 老板说：“Harness 和 Context、Prompt Engineering 不一回事吗？” 我说：“你可别掺和了”
点
分
享
点
收
藏
点
点
赞
点在看