Founder, Digital Passion. Non-dev building with Claude, ChatGPT and GrokBot. Daily takes on AI agent news and real tests, for builders who run multiple agents.
GPT 6.1 Terra came out today, so I tested one thing: does extra-high thinking make better images?
3 real image jobs, 5 model setups, identical prompts (hash-checked), one shot each.
Short answer: no. It cost more and bought nothing. Here's the setup.
@stretchcloud This is why I don't let my agents settle things by agreeing with each other. One of them checks what's actually true in the issues and the code, and that wins.
@MrTacticalX $5 vs $58 is a huge gap. What I'd want to see next is how many rounds of fixes each one needed before it was actually done, since that's where most of my bill goes.
@TokenGremlin I’m hopeful. We need the quality and efficient combo to return so it’s on par. Or else it’s not going to look good. And what surprises do you think they preparing for the IPO?
@MrTacticalX Almost the same bill even at twice the token price is the interesting part. What matters to me is the cost per finished job, not per token.
@stas_sorokin_@Grirku Mostly my morning run: reading through the day's posts and my notes to plan what to work on. Not that big, so one model doing it all has been fine for me.
@Voxyz_ai Good list. Have you tried setting these as the output style in settings instead of adding more lines to the instructions? Curious if you saw any difference in results.
@rileybrown The part I like is reviewing a finished link after the agent kept working at home. That's how I want my day to go: I check in, it continues, and I look at the result when I'm ready.
@stretchcloud Having a different model check the work makes sense to me. I'd rather get a second opinion than have the same model mark its own homework.
@Joshua_WD The dreaming pass is the part that clicked. Wrong notes that stick are worse than no notes. Keeping memory thin and cleaning it later is what I want too.
@AlexFinn Been running Grok Bot the same way. The stuff that sticks is the quiet checks that run while I'm away, then wait for my yes before anything goes out.
@Verdana86@NotionHQ My agents don't rely on one memory source. They each have their own memory, plus Linear for issues, Obsidian as a second brain, and GitHub for code. For me, logging things properly matters more than memorizing everything.