Nobody pays for a score
Affiliate links help pay for the testing — but they never touch the ranking. Every tool on the board earns its place on the published rubric, and the list is not for sale at any price.
Independent testing · re-run every 30 days
We ran 8 AI coding agents through the same 12 tasks.
No vendor writes a word here. Same repo, same prompts, same reviewer, scored on a published rubric — and re-tested when the models change.
8 tools tested · 18 task runs · Last update: 22 Jul 2026
Illustrative dataThe board · AI coding agents
Ranked by tested score, never by payment.
Illustrative data| # | Tool | Score | From | One-line verdict | Verdict link |
|---|---|---|---|---|---|
| 01 | Cursor | 8.5 | $20/mo | The best all-round coding agent we have tested — provided you are willing to live inside its editor and watch the bill. | Verdict → |
| 02 | Claude Code | 8.5 | $20/mo | The one to beat on long, cross-file refactors — it holds context where the others lose the thread. | Verdict → |
| 03 | GitHub Copilot | 7.8 | $10/user/mo | The safe team default — cheap per seat, everywhere your engineers already are, if you do not need a full agent. | Verdict → |
| 04 | Windsurf | 7.4 | $19/mo | Promising, but it regressed on the monorepo suite in July — recheck in August before you commit. | Verdict → |
| 05 | Cline | 7.1 | Free + API | The budget pick — a free client on your own API keys, so you pay for tokens and nothing else. | Verdict → |
| 06 | Aider | 6.9 | Free + API | Capable in the terminal, but the workflow asks too much of the user to make the shortlist. | Verdict → |
| 07 | Replit Agent | 6.4 | From $20/mo | Great at spinning up greenfield projects, weaker the moment it meets an existing codebase. | Verdict → |
| 08 | Devin | 5.8 | From $500/mo | Ambitious autonomy that does not yet pay off — too little finished for the price to make the shortlist. | Verdict → |
Every verdict has a date
A score with no date is marketing. Each verdict here carries the version we tested and the day we ran it — and we re-test when the models change, not when a vendor asks.
Three picks, never thirty
Most directories list everything and recommend nothing. We name 3 picks — best overall, best on a budget, best for teams — and cut the rest with a reason and a date.