Loading the page.
Topic hub
index/AI agents
updated 23 September 2026AI coding agents
Programs that read a repository, edit files, run tests, and take a task through to a commit on their own. The whole landscape - on one map: tile size - strength in our ranking, color - vendor camp.
Size = our score · click - profileOpenAIAnthropicGoogleCognitionCursor
players in the field
18
CLI trio installs / week
33.1M
week over week
-22%
holding up best
Gemini CLI -5%
live race of the week - analysis below
Race of the week
Who's driving the market right now
CLI agents were downloaded 33.1M times this week - 22% less than last week. Codex is confidently first (20.6M/week vs. 12.1M/week for Claude Code); Gemini CLI - a distant third (406K/week).
Claude Code12.1M/week
Codex20.6M/week
Gemini CLI406K/week
api.npmjs.org · data points as of 23.09 · updated hourly · CLI trio only: the rest use other distribution channels
how the CLI trio's week is split (npm)
Claude Code37%
Codex62%
Gemini CLI1%
2.0
Our score is the tool’s overall strength: benchmarks, development activity, ecosystem. It's not the same as popularity: compare it with the race above.
Codex
97.0
Claude Code
91.0
Cursor
90.0
Devin
78.0
Windsurf
77.0
Copilot Agent
69.0
Jules
65.0
Antigravity
64.0
3.0
Catalog: 18 agents
A profile is your entry point: prices, limits, benchmarks, track record, and access from Russia, all with verification dates.
Codex
97.0OpenAI
OpenAI coding agent for code and computer work: desktop app, CLI (open source), IDE, web/cloud, GitHub review, SDK, and phone control; primary model - GPT-5.6 (Sol/Terra/Luna family, since 09.07).
Claude Code
91.0Anthropic
Anthropic AI agent for repository-scale tasks: a terminal-first process with auto mode, dynamic workflows, subagents, background agents, and control through diff, tests, limits, and usage.
Cursor
90.0Cursor
Cursor - an editor for AI development: best where work happens with eyes and hands in the IDE, not in a separate cloud runner.
Devin
78.0Cognition
Devin - a cloud contractor for tasks: give it an issue, get a PR attempt, and review it as work from an external contributor.
Windsurf
77.0Cognition
Windsurf - an IDE agent at the intersection of the editor and a managed workflow: assess it with fresh eyes, not memories of old Codeium.
Copilot Agent
69.0GitHub
GitHub Copilot Coding Agent - an executor inside GitHub: turns an issue into a branch/PR if the team knows how to write issues.
Jules
65.0Google Jules - an async runner for GitHub tasks: yes for small assignments, no for a full workstation.
Antigravity
64.0Google Antigravity - an experimental agent-first IDE: the direction is interesting, but maturity must be proven on a branch.
Replit Agent
51.0Replit
Web-based app builder within Replit: suitable for fast prototypes and beginners, but does not replace a serious local coding workflow.
Lovable
43.0Lovable
Lovable - a fast builder for demos and MVP screens: it sells speed to product, not engineering maturity.
Aider
40.0Aider
Aider - an honest open-source CLI: less polish, more control; you choose and pay for the model yourself.
Bolt
36.0StackBlitz
Bolt - a browser-first builder: quickly create a web prototype, then move it into a proper engineering process.
Junie
32.0JetBrains
JetBrains Junie - an AI agent for JetBrains teams: if the IDE is already chosen, test it; if not, this is not the first decision point.
Augment
28.0Augment
Augment Code - an enterprise agent for large context: needed where the repository is already bigger than one developer's head.
OpenCode
20.0SST
OpenCode - an open agent harness: for those who want provider control and do not want a closed default.
Cline
13.0Cline
Cline - an open IDE agent with approvals and checkpoints: control is good, but responsibility for configuration remains with the user.
ChatGPT
—OpenAI
A universal OpenAI chat assistant on GPT-5.6 (Luna/Terra/Sol): chat, voice, images, and ChatGPT Work agent mode, which takes a task through to a finished artifact.
Gemini
—Google's proprietary AI assistant powered by Gemini models (3.8 Flash and 3.1 Pro): three models without a subscription, up to 1M-token context on Pro/Ultra, Deep Research, voice, and Omni Flash video; built into Search, Workspace, Android, and Chrome.
4.0
Pairs: agent + model
An agent’s score is always an “agent + model” pair: Codex is measured with GPT-5.6, Claude Code - with Fable 5. Compare pairs, not brands.
← Claude Code + Opus 4.8Codex + GPT-5.6 →
Terminal-Bench 2.106.2026
78.9%
67%
Rope tilt = who is stronger on this benchmark; bright side - winner
5.0
What's new in the field
Events from the profile track records of two flagships: releases and fixes push up, incidents and limit policy push down.
2026-08-13
Codex: IBM embeds Codex, GPT-5.6 and ChatGPT Work in IBM Consulting Advantage
2026-08-11
Codex: Linux app in preview + import settings and chats from Claude Code and Cursor
2026-08-09
Codex: Mass limit resets on August 9, 10, and 13 - amid demand for GPT-5.6 Sol and the 15 million user milestone
2026-08-07
Codex: CLI 0.147.0: portable Agent Plugins, --approve-for-me flag, MCP protocol 2026-07-28, skill migration from Cursor
2026-07-30
Codex: GPT-5.6 API price cuts: Luna −80%, Terra −20%; Sol unchanged
6.0
How to choose for yourself
Don't choose the “best overall” - choose for the job: the 2026 market runs on two-agent setups.
Complex work in your own repo
Claude Code →
Affordable pipeline for repetitive tasks
Codex →
AI-IDE with autocomplete
Cursor →
Open client with your own key
Cline →
Which AI coding agent should you choose in 2026?
Market split: Claude Code - depth on a complex codebase, Codex - an affordable parallel assembly line, Cursor - an AI-IDE with autocomplete. The pro standard - a two-agent setup: tasks are routed by type.
Are there free agents?
Codex (in ChatGPT plans, including Free) and open-source clients like Cline with your own API key offer free access. Claude Code has no free plan - entry starts at $17/mo.
Do agents work from Russia?
Officially - no: Anthropic and OpenAI do not support Russia, and payment with Russian cards is impossible. Viable routes - a foreign card or APIs via aggregators in rubles (~1.5–2×); each profile includes an “access from Russia” section.
Agent and model - what's the difference?
A model - the brain (GPT-5.6 Sol, Fable 5); an agent - the hands: it reads the repository, edits files, runs tests, commits. An agent's score is always measured as an “agent + model” pair - compare pairs, not brands.
7.0
Where to go next
The hub - an entry point into the topic; depth lives in profiles, concepts, and practice.
Try it yourself
- Guides and practice
- MCP server catalog for agents
- Token maximization: get the most from subscriptions