OD Chenko

Qoduck, the free mystery model in Qoder, fixed real bugs in 15 of 21 tries for $0. Nobody will say who made it. People are guessing Qwen 4. 7 bugs, 3 tries each, judged by the maintainers' own tests. The answers came in three kinds: → exactly right. marked (Markdown to HTML) crashed when a plugin was named "constructor". The model wrote the same one-line fix the maintainers merged → right idea, new side effect. axios timed requests with the system clock, which can break a Next.js build. The model switched to a stopwatch timer, but that timer returns fractions, and axios reports whole milliseconds → confidently wrong. fastify rejected every request when its header rules came from a shared schema. Twice the model explained the cause in detail, and the request from the bug report was still rejected So I added one line to the fastify prompt: turn the bug report's example into a test and walk it through your fix. It fixed the reported bug 3 times out of 3, up from 1. And all 3 times it broke the case right next to it, the one I never mentioned.

One click in Claude Code and my test password was one commit away from GitHub. The Claude Code team says leaked files like this were added to git by hand. I never touched mine. The click was "Yes, and don't ask again". It saves the command you approve into a project file, passwords included. Apiiro found 1,122 of these files with real secrets on public GitHub. what I did: → a curl with a test password, approved with "don't ask again" → Claude Code saved it to .claude/settings.local.json → the folder wasn't a repo yet, so no auto gitignore → git init, git add . and the password was staged same click in a folder that was already a repo: → the password still went into the file → Claude Code added the file to my global gitignore, git skipped it same password in .env: → Claude Code used $STAGING_PASSWORD instead → the password never showed up in the command → "don't ask again" wasn't even offered In my runs the leak needed two things: a password typed into the chat, and a folder that became a repo later

On tasks a cheap model already failed, the next model succeeds about half as often as it usually does. My old routing was built on exactly that handoff. That's from Route Anything, a new NUS and Berkeley study with 374,401 scored runs. The drop showed up in all 42 model pairs they tested. Half is the median. What paid off was sorting tasks before the first call. On their coding benchmark that router never once called Opus 4.6, got 25.5x Opus's score per dollar and lost under half a point. Sending everything to the cheapest model, Qwen3-Max, got 20.1x but dropped 4.2 points. And on wall-clock time, plain Opus beat every cheap-model-plus-checker chain. It simply answered faster. Paper ↓

The Claude Code tip "Sonnet writes, Haiku searches, Opus advises" does save 60%. But only without the Opus part. With Opus advising, my bill went up 2.2x. Same bug, 3 runs per setup, every fix passed tests the model never saw. Cost of one fix: - default, Opus does everything: $0.58, 89 s - the tip with the usual settings: $0.41, 164 s. Haiku never searched, Opus never advised. Just Sonnet - all three models actually working: $1.30, 389 s. Opus advising alone was $0.78 of it - Sonnet writes, Haiku searches, no Opus: $0.23, 154 s The cheapest setup wasn't the fastest. Opus alone finished in 89 s. What finally got Haiku searching, for 2-3 cents a fix: CLAUDE_CODE_SUBAGENT_MODEL=haiku CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 claude --model sonnet plus one line in the prompt: "use the Explore subagent for searching". Without FORCE, the Haiku setting doesn't reach the built-in search agent. Setup details and why the advisor costs so much ↓

I tried to get an AI to open a password file it was told to leave alone. 50 runs, it never did. Then I gave it a job that needed the password, and one of the two models opened it.

A perfect transcript can still miss the whole call. Nemotron 3.5 Transcribe just hit #1 on Voice Arena, 32.43% cpWER and the fewest dropped words on the board. Reading the words is basically solved. The voice agents I saw rarely broke on a wrong word. They broke on tone. Someone says "that's fine" in a voice that means the opposite, the agent only gets the text, and it believes them. A better transcriber doesn't fix that. It just transcribes the flat version faster. Models that hear the mood exist now, but they cost you the clean transcript most teams still build on.