This is crazy. OpenAI's cheapest model just beat Fable 5 on my bug bench. Better and cheaper. 33 bugs fixed against 24. $1.80 against $68. That's GPT-5.6 Luna at max reasoning effort, run through Bug…

63 of the 105 bugs survived every model. The diff is ground truth, not the model's own report. Blind judges, anonymized submissions, a withheld answer key. One run each, so precision is about one fix…

Opus 5 fixed 11 of the 45 bugs I hid in my own repo. Opus 4.8 fixed 2. Same repo, same prompt, same setup. Opus 5 just shipped. I ran it straight through Bug Hunt Bench: 45 bugs across my VS Code…

Everyone's citing an 80% cut to Claude Code's system prompt. I captured the real prompts, on every model. Three things the headline skips: 1. It's closer to 70%. The 80% is the memory-off count:…

Full method, the xhigh CLI trick, and my delegation setup are in the Fable 5 guide:

I spent ~$45 testing Sakana's Fugu Ultra this week. Behind a VPN (it's geo-blocked), 10 runs. The pitch is interesting: a multi-agent system you call like a single model. Let me save you some tokens:…