Last reviewed October 3, 2026
What's running right now
Everyone asks which model I use. Wrong question. I use a different one for every job, and they have to earn it.
Here's how. Every job my AI does, I capture the real prompt it sends. Not a toy test. The actual morning prompt, with the actual news in it. Then I hand that same prompt to thirteen models. Nobody knows whose answer is whose.
Code grades the boring stuff first. Did it follow the format? Did it make up a link? Is that quote really in the transcript? Then two judges from different labs score the rest.
One rule. A judge never grades its own lab's homework. I tested it. Each judge scores its own family about two points higher than the other judge does. So I only count the judge from the other side.
The lineup
News digests and podcast summaries
GPT-6.1 Sol
Took over from GPT-5.6 Sol on October 1. Clear wins on scoring podcast episodes (8.3 vs 6.3) and the weekly podcast digest (8.0 vs 6.0). Here's the part that surprised me. Back in September, one health digest came back with four links that weren't in the source. I added one guardrail to the prompt and it dropped to zero. So did every other model. The model wasn't the problem. My prompt was.
Research briefs and LinkedIn post ideas
GPT-6.1 Sol
I moved this off Grok 4.7 in September, then up to 6.1 Sol in October. It scores 8.1 to Grok's 7.0, and it's done in under two minutes. Grok took five.
My morning and evening council
GPT-6.1 Sol + Grok 4.7
Eight AI execs argue about my day, then a chief of staff settles it. 6.1 Sol writes the exec briefs now. Cleaner than Grok's across 72 tests, 8.8 to 8.1. Grok still runs the chief of staff, because I can't score that step fairly yet.
Writing in my voice, and the auditor that checks my AI's work
Next up
Not tested yet. When they are, they land here.
What I don't trust yet
Some jobs still have one or two test cases. That's an anecdote, not a verdict. My two judges pick the same winner 51 times out of 142. That's why the code checks carry the weight, not the judges.
When a new model ships, my system texts me. It asks if it should run the evals. I say yes or no. Nothing switches on its own. I update this page by hand when something wins.