I heavily coded with three top AI models for the last two weeks: GPT 5.6 Sol, Fable 5, and Qwen 3.8 Max.
I made them review each other. The results are quite unexpected.
Fable 5 > Qwen 3.8 Max > GPT 5.6 Sol https://t.co/VoRKTHCVjL
I should clarify that the volume of the code is not very representative in this context.
Agents were given with different tasks.
But intuitively, I still feel like Sol produces too much code, unnecessary double-checks and guardrails.
I heavily coded with three top AI models for the last two weeks: GPT 5.6 Sol, Fable 5, and Qwen 3.8 Max.
I made them review each other. The results are quite unexpected.
Fable 5 > Qwen 3.8 Max > GPT 5.6 Sol https://t.co/VoRKTHCVjL
Qwen 3.8 Max, running via Claude Code in "ultracode" mode, commented on research previously generated by the same Qwen 3.8 Max model but via Qwen Code:
"The direction is correct (~70% alignment with my analysis), but the implementation is dangerous and technically unsound in half the points."
The difference is striking. Claude's harness really makes the Chinese model work hard, exactly as it should.
In contrast, Qwen Code doesn't even allow you to set an "effort" level.
Working hard on a fully automated go-to-market tool for solo founders - WelderGTM.
10 socials, text and video posts, full autopilot.
Full means really full - from exploration to generation and posting.
Coming soon, join the waitlist to sucure early access.
@CountlyTracker app got really positive feedback on Reddit.
Builders, create positive and helpful content related to your products, and Reddit will pay off.
Overall, Countly App is doing really well. 10% conversion from download to paid. 20% from download to trial.
Check the Countly app. It's a very helpful thing for all travellers and immigrants. I'd love to hear your feedback