Roger Kaleba

Product designer for AI, fintech & web3 founders | I design the moments users decide to trust you, pay, or leave

[Roger Kaleba @RogerUiUx]

Users don't leave bad products. They leave unclear ones

[Roger Kaleba @RogerUiUx]

Multiplay brand identity.
One mark, one indigo, and a wordmark nobody's allowed to retype

[Roger Kaleba @RogerUiUx]

@maazinsf @ycombinator 97% never inspected is wild so the whole premium basically rides on whatever the owner typed into a form

[OpenAI @OpenAI]

GPT-6.1 Sol: near-Astra intelligence for a fifth of the price.

It’s the most cost-efficient model for its performance available today.

[Roger Kaleba @RogerUiUx]

@AmenaiSabuwala @claudeai everyone's asking skip what, but i read it as trying @claudeai without an account first, @ChatGPT already lets you do that, so it's a bigger change than the background tbh. @AmenaiSabuwala was that the idea?

[Roger Kaleba @RogerUiUx]

@tj_littlejohn nice, being able to watch it is what would make me actually hand it my phone

[Roger Kaleba @RogerUiUx]

@wuweiweiwu 2 years making tests easier to write and then launching the thing that means you never touch them again. respect for eating your own product lol

[Roger Kaleba @RogerUiUx]

@ParallelAILabs Thanks. The case I'm still working out is an agent editing on someone's behalf, is that a square, a circle, or both? Feels like "who did it" and "who asked for it" both matter there

[Roger Kaleba @RogerUiUx]

@DARKKEVN yeah exactly, it hands you the list of what's still yours to check. an agent that just says "done" makes you redo the whole review yourself

[Roger Kaleba @RogerUiUx]

The most interesting part of the GPT-6 Sol launch isn't the price cut.

It's one reply, shown before and after. The old one says it's done. The new one says what it changed and what it checked: desktop, narrow mobile, the back button.

OpenAI called that the better answer.

Agents are learning what good products already knew. Telling people what you checked is part of the job.

OpenAI: APIs and tools for building AI products by @sama https://t.co/8Pvls95ntu

[Roger Kaleba @RogerUiUx]

When people and AI agents work on the same doc, "who did this?" can't be a guess.

Been designing a brand where the answer is a shape. People are circles. Agents are rounded squares. The mark is both, side by side, never merged.

Full concept on Behance soon

[Roger Kaleba @RogerUiUx]

Three models tied at 97%.

So the model isn't the difference anymore. The 3% is. What your product does when the answer's wrong is the part users actually remember, and it's the one part you can't buy from a lab

[Guillermo Rauch @rauchg]

We ran fresh Next.js evals. The tally:
① Opus 5.5 [𝟿𝟽%]
② GPT 6 Sol [𝟿𝟽%]
③ Fable 5.1 [𝟿𝟽%]
④ Grok 4.7 [𝟿𝟺%]

Notably, Grok is 2x-7x cheaper

[Roger Kaleba @RogerUiUx]

@ClaudeDevs Putting cost per task in /usage instead of a pricing page is the right call, I think. One layer I'd add for plan users: how much of the 5h window a task eats. A lot of the replies here think in windows, not dollars.

[Roger Kaleba @RogerUiUx]

@ycombinator What doesn't carry over from coding agents is undo. A bad code step is a revert, a bad robot step is something on the floor. So I think the approval moment has to move from after the action to before it

[Roger Kaleba @RogerUiUx]

@simonw Simon's full writeup, pelicans included: https://t.co/gPgTpY0NMv

[Roger Kaleba @RogerUiUx]

An AI spent 20 minutes thinking about a pelican on a bicycle this week, then said nothing.

@simonw gives every new model the same test. On Tuesday he gave it to Claude Opus 5.5 on its highest thinking setting. It planned the beak, checked the pelican's shin length, planned a fish in the basket, then ran out of room before drawing anything. Twice. Nearly 20 minutes and $2.56 each time.

Same day, the price war started. Opus 5.5 got 20% cheaper with cached input down 60%, and OpenAI shipped GPT-6 Sol and Luna at half the price of what they replace. Long agent runs just got a lot cheaper.

So I think products are about to run longer, and more people will sit through their own version of those 20 minutes.

That's the screen most products don't design. What's it doing, how far along is it, what's it costing me, can I stop it, what do I get if it fails?

Simon got a blog post out of his 20 minutes. Your users probably just get a spinner

[Roger Kaleba @RogerUiUx]

@uriel_builds @adriankuleszo yeah that makes sense for now

[Roger Kaleba @RogerUiUx]

Dashboard explorations for an AI money app

[Roger Kaleba @RogerUiUx]

@onethomaschan @ycombinator #1 has a quiet bonus i think. the launch post ends up being the clearest pitch you'll write, because you wrote it before polishing. worth watching which lines people quote back and moving those onto the site