Why similar AI benchmark scores can hide different models
DeepSWE puts Luna Max 2.2 points behind Sol High at roughly one-sixth the attempt cost. I explain why the models can still feel far apart in repository work.
Hands-on notes on AI and software development
Page 3 of 3
My coding agents forget the repo. I built them a map
I built a four-layer context management system that routes coding agents to current facts. It works, but stale documentation is still the hard part.
GPT-5.6 vs Claude 5: why the benchmarks disagree
I compared Sol, Terra, Opus 5 and Fable 5 across coding benchmarks. The winner changes with the task, effort setting, agent setup and budget.
Kimi K3 matches GPT-5.6. It takes 4.6 times as long
Kimi K3 ties GPT-5.6 medium but takes 4.6 times as long. GLM-5.2 is cheap per token yet costly per task. I checked where both models still win.
I compared eight Convex alternatives. I'm not switching
Appwrite and InsForge came closest, but one TypeScript backend still makes Convex easier for my AI agents. EU hosting is the expensive catch.
GPT-5.6 Sol over-engineers. I still use it
In my release audit, GPT-5.6 Sol found nearly six times as many possible issues as Fable 5. Most failed triage, so I changed how I use it on coding tasks.
Opus 5 is a great employee but a frustrating boss
Opus 5 matches Fable 5 on benchmarks at half the token price, yet it can be painful to supervise. Here is where Fable still earns its place.