The job: Make real changes in an existing codebase: features, refactors, bug fixes and tests, with you reviewing.
For: Developers and vibe coders working in a real repo
Anthropic's agentic coding tool for the terminal, IDE, desktop and web.
The terminal-native agent many builders reach for first; maxed builder-fit (CLI, MCP, hooks). Heavy use pushes you to the $100-200 Max plans.
OpenAI's coding agent across CLI, IDE, desktop app and cloud tasks.
Buffed in Sep 2026 with GPT-6 Astra, and included in every ChatGPT plan. Plus-tier Astra limits are tight.
AI-first code editor with agents, cloud agents and Bugbot code review.
Slickest editor UX. Pricing reshuffled on Oct 12, 2026 into The X Subscription; third-party models now billed at API list price.
GitHub's AI pair programmer with a coding agent, CLI and cloud agent.
Cheapest paid entry ($10/mo) with unlimited completions and a GitHub-native cloud agent. Agent use now burns AI credits.
Open-source autonomous coding agent for VS Code and other editors. Bring your own model.
Open source, bring any model (or run local). Max control; cost depends on the model you choose.
Google's open-source terminal coding agent powered by Gemini.
Best free tier on the board, but nerfed to Flash-only for free users on Mar 25, 2026.
Cognition's agentic IDE, renamed from Windsurf, with local and Devin Cloud agents.
Windsurf, renamed in June 2026, now a hub for local and cloud agents. Quota overages billed at API prices.
AWS-backed agentic IDE built around spec-driven development.
Spec-first workflow is a distinct playstyle. Clear credit tiers, small free allowance.
Open-source AI pair programming in your terminal, git-native.
Free, open source, git-native classic. Release cadence has slowed (last stable Feb 2026).
How to read this list: job score out of 100 using our public rubric. Tiers: S ≥ 85, A 75-84, B 65-74, C 50-64, D < 50.
This list is about working in a codebase that already exists, not generating a new app from a prompt (that’s the vibe coding MVP list).
Hands-on test pending. Rex’s test for this list will run each agent on the same open-source repo with a fixed task pack (add a feature with tests, fix a seeded bug, do a cross-file refactor), recording pass/fail, iterations, wall-clock time and cost. Public benchmarks such as SWE-Bench Pro will be cited alongside, not copied.