~ / best

Best AI Coding Agents, Models, and Stacks

Start with your workflow: terminal, VS Code, open source, enterprise, low cost, long context, or full agent plus model stack.

Best CLI AI Coding Agents

Terminal-first coding agents are best when you want repo-level edits, scriptable workflows, and less IDE lock-in.

11 ranked entries

Best Open Source AI Coding Agents

Open-source agents are strongest when you want provider choice, local-model support, and transparent agent behavior.

9 ranked entries

Best AI Coding Agents for VS Code

IDE-focused agents are best for daily coding, codebase navigation, completions, and fast multi-file edits.

7 ranked entries

Best Browser AI Coding Agents

Browser coding agents are useful when you want to build, preview, and ship without managing a local development environment.

3 ranked entries

Best Enterprise AI Coding Agents

Enterprise teams usually prioritize governance, familiar integrations, security posture, and support more than raw benchmark score.

7 ranked entries

Lowest Cost AI Coding Stacks

These stacks favor free agents and low-cost coding models while preserving enough benchmark performance for practical work.

62 ranked entries

Best Open Source AI Coding Stacks

Fully open-source stacks trade some convenience for control, portability, and lower vendor risk.

26 ranked entries

Highest Score AI Coding Stacks

Current models paired with an agent that runs them, then the June 2026 stack snapshot. ARscore has no formula behind it, so the table is ordered by the model’s measured DeepSWE score.

12 current pairs, 63 archived

Best Long Context Coding Models

Long-context models are valuable for agent workflows that need large repositories, logs, documentation, and multi-file reasoning.

36 ranked entries

Best Open Source Coding Models for Agents

Open models pair well with open-source agents when you need cost control, portability, or local deployment options.

13 ranked entries

Current models, paired with an agent that runs them

Ordered by the model's DeepSWE 1.1 score. No stack-level number here, because none of these pairs has been measured end to end.

#ModelDeepSWEAgentAgent archContextCostOpenPairing basis
1Gemini 4 Argon
Google
77.9Gemini CLI481Mnot listedNovendor harness
2GPT-6.1 Sol
OpenAI
75.2ChatGPT Codex581.1M$2/$10Novendor harness
3Claude Opus 5.5
Anthropic
74.2Claude Code661M$4/$20Novendor harness
4DeepSeek V4.1 Flash
DeepSeek
74.2Aider43not listed$0.15/$0.60Yesopen weights, open agent
5GPT-6 Astra
OpenAI
74.1ChatGPT Codex58272K+$10/$50Novendor harness
6Gemini 3.8 Flash
Google
73.7Gemini CLI481M$0.75/$3.75Novendor harness
7MiMo-V2.6-Pro
Xiaomi
71.9Cursor561M$0.435/$0.87Yesopen weights, IDE agent
8Claude Sonnet 5.5
Anthropic
71.0Claude Code661M$2/$10Novendor harness
9Claude Fable 5
Anthropic
70.0Claude Code661M$10/$50Novendor harness
10GLM-5.3
Z.ai
66.9OpenCode48not listednot listedYesopen weights, open agent
11Mistral Large 4
Mistral AI
61.7Aider431M$0.68/$2.09Noopen weights, open agent
12Qwen3.8-Flash-Next
Alibaba
58.7Cline47not listednot listedYesopen weights, open agent
Why no ARscore above. ARscore was a stack-level figure measured in June 2026 against a model set that no longer contains a single current leader, and there is no formula behind it — so it cannot be recomputed on today's models. The table shows the two layers separately and orders by the one that was measured. The full June snapshot has all 63 stacks.