Dan's hands-on AI testing lab

AI tools, tested by a working developer.

I use AI coding agents to build my own games and tools. Here I share the results, the bills and what went wrong, so you can decide which tools are worth trying.

Open the tier list All test reports The tests on YouTube

9 test reports
60 receipts
12 personally ranked agents

The latest verdict, as its receipt

DanManREAL · AI lab

Test receipt

Field testOct 21, 2026

Gemini 4 Argon

Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistency

Clips reviewed: 14 MP4s
Long-form segments: 12 of 12 PASS
Short 1: ISSUES

+ 1 more receipt in the report

Verdict

Useful review, inconsistent grades

Tier list One tool on real work, not a controlled comparison against other agents.

danmanreal.com/tests/gemini-4-argon-video-review

DanManREAL · AI lab

Test receipt

Head-to-headOct 19, 2026

DeepSeek V4.1 Flash

Does an extra AI review step pay off? A small DeepSeek trial said no

Task 1 time: direct vs extra review: 182.43 s vs 313.91 s
Task 1 token cost: $0.01541 vs $0.03902
Task 2 time: 175.14 s vs 188.06 s

+ 4 more receipts in the report

Verdict

Extra review cost more on both repairs

Tier list Two workflows on the same model, so it ranks no agent against another.

danmanreal.com/tests/deepseek-direct-vs-extra-review

DanManREAL · AI lab

Test receipt

Field testOct 23, 2026

ChatGPT

ChatGPT + Codex built 33 new DMR_REMOTE tools

New tools with successful real calls: 33
DMR_REMOTE tool catalog: 45 → 82
The original ChatGPT agent hunt: About 25 min

+ 1 more receipt in the report

Verdict

33 new working tools. One session.

Tier list This records a working collaboration in my workshop; it does not rank models against each other.

danmanreal.com/tests/dmr-remote-33-new-tools

Replaying published report 1 of 9

Read this report

Latest verdicts

Hover one to print itAll 9 reports

My rankings · saved 2026-10-23

AI Coding Agents Tier List

12 of 12 personally ranked from daily use. 8 have published reports. These are my judgments, not controlled benchmark standings.

Claude Opus 5.5, 1 receipt
Claude Sonnet 5.5, not yet tested
Claude Haiku 5.5, not yet tested
DeepSeek V4.1 Flash, 4 receipts
GPT-6.1 Sol, 1 receipt
GPT-6 Astra, not yet tested
GPT-6 Luna, not yet tested
Gemini 4 Argon, 3 receipts
GLM-5.3, 1 receipt
Sonus, 1 receipt
Qwen 3.8 Max, 1 receipt
Muse Wonder Full 1.0, 1 receipt

Open the tier list