Pany Haritatos

With Argon coming out, I've been using Gemini Flash 3.8 less and less in order to save quota. I have been reaching for open source models in an attempt to get max intelligence on a reasonable budget. To give you a sense of scale, I'm at 83B (with a B) tokens in the last 30 days on my solo projects. I've spent a ton of time with MiMo v2.6, DeepSeek 1017 and the GLM models. I think my favorite was Mimo even though it's a bit slow. Everyone seems to love DeepSeek 1017, but it's deceptively token-hungry, and way too confident which is a baaaad combo. I'm breaking up with all of them unfortunately. I was looking at the live data of our game making users on run.world and pretty consistently Haiku was kicking ass in terms of quality per dollar. I've been using it a ton this week and I can see why. It's surprisingly cost effective at API rates - even more so than many of the open source models, and it is mostly better on the vectors that are important to me (not always, but usually). So, for now, Haiku 5.5 will be my daily driver while I save my subs for calling in airstrikes with supermodels. For "hard work," Opus 5.5 High is still the boss. I reach for it for anything tough, super R&D heavy or requiring data analysis. Kimi 3.1 Pro shows a lot of promise and has surprisingly good taste when it comes to game design & art direction. Sol 6.1 continues to drive me absolutely crazy for a bunch of reasons, but.. it's still in the rotation even though I'm mad about it.

You can run the new 1017 Deepseek at 1500 toks/s on Lithos, and it has ZDR. I'm going to try it out as my main driver this week. Wish me luck!

Posting from the front lines, here's my latest hot take on model use in development (especially games), which has changed a LOT in the last month: * Opus 5.5: Super smart and reliable, though a little slow and weirdly pedantic in an annoying way. Also, it feels like it is getting a nerf? I love how token efficient it is, though. * Sol 6.1: Maddeningly slow. But as a Codex model, it's the GOAT for complex, long-running tasks. It just doesn't give up on crunchy, complex tasks. * Grok 4.7: Hot garbage. I will be canceling my sub. * Mimo 2.6 Pro: Easily the most unhinged model I've worked with. Not sure if it's the provider or the model itself, but it goes completely off the rails in all my tests. I'm disinclined to keep trying given that GLM 5.3 seems better basically in all respects. * RIP Astra. It was a wild and lovely affair, over too soon. Day to day, I'm working primarily with Gemini Flash 3.8 for the speed and heavily subsidized tokens, reserving the "supermodels" for planning, audits, and visual work. You have to teach the lesser models to call in supermodel airstrikes, but once you do, you're working at the pareto frontier for a fraction of the price. Coming up... I'm really interested in Gemini Flash 4.0. I'm also experimenting with GLM 5.3 on the 4x providers to see if it can displace Gemini. GLM 5.3 is fast, relatively cheap, and beats GF3.8 in most tests, but open-source models have many sharp edges, so we'll see.

And in minutes you can also.. publish to mobile, monetize with ads or even charge users for upgrades.

I never thought I'd be posting this. Grok 4.6 has become the model I reach for most. It's not the smartest thing I can call. It's smart enough, and that turns out to be different and more useful. Speed is most of it. If a model is marginally better at my daily work but takes 3x as long, that's not a trade I'm making any more. The rest is generosity. My other subs get stingier exactly as I get more dependent on them. Grok doesn't. I can lean on it without feeling fleeced. When something's novel or genuinely hard I still go to the heavies. That's 10% of my week, not 90%. xAI is on the right track.