Matias Villagrán

Today we’re releasing DeepSWE, a new standard for agentic coding benchmarks. On public leaderboards, top models often look relatively close in capability. DeepSWE shows where they actually diverge, reflecting the realistic experience of developers in their day-to-day work.

getting closer and closer to the UI I wanted for my roguelike card game still kind of crazy to me that this is 100% Swift/SwiftUI without the Codex app, I don’t think I would’ve iterated anywhere near this fast

@theo 100% codex UI

322 apps already deployed on my cloud 👀

@romainhuet @eeismann please pay attention to LATAM companies too. Pushing Codex almost alone here 😅

@eeismann Is there anything we could do to help you use Codex at your company?

Feeling left out because my company doesn't use Codex. C'mon Anthropic - let's innovate on the UX of your desktop app.

@RoFerreiraDev 🫠

@Zickox Jaja clásico. El bug más raro en el peor momento posible.

codex crashed so hard it started speaking in question marks 😅

@trikcode Skill issue

Claude code is still better than Codex Prove me wrong

@Dimillian so… what’s the roadmap for iOS 27? 👀

Is this a roadmap

No me considero engineer 100x pero la semana pasada apoyé a 4 equipos técnicos que trabajan con 4 tecnologías diferentes sacando issues en producción: 2 fixes a backend escrito en GO 1 feature a backend escrito en NestJS 1 fix en Android escrito en Kotlin 1 feature en iOS escrito en Swift Algo impensado hace un tiempo atrás. Por cierto no fue en un proyecto pequeño, fue un proyecto con más de 2 millones de sesiones de usuarios mensuales

Yo ya estoy evidenciando ingenieros de software que no son 10x engineers sino 100x engineers. Usan varios agentes de desarrollo de software, tiran decenas de sesiones a trabajar al mismo tiempo. Hacen el trabajo de meses en días. Una locura.

@yeguacelestial $30k usd por acá 😬

quedan diez dias para el subsidio de tokens de OpenAI. en un mes, he gastado $4.2k en USD de tokens solo usando el plan de $100 USD. una puta locura lo de los subsidios, si me preguntan. junio será ESENCIAL para ver qué tan efectivo es GPT 5.5 para hacer producto sin tokens pagados.

@freddier also, OpenAI Codex is capable of working for 101 hours or more 👀

OpenAI Codex is capable of clicking all buttons in an app it develops, check it the behavior works as expected, find the bugs and fix them... if you explicitly ask for that. It also has an in-app browser where you can mark changes, if your app is a website. Very cool work. The next version of Claude Code is going to be insane.

@qi9098 100% agree! Great summary 🔥

101 hours is the interesting part. /goal starts to make sense when Codex can keep reading tests, screenshots, and visual diff feedback over time. Appshots matters for that too because the thread gets real app/window context, instead of a hand-written UI description. I mapped the rest of the Codex Thursday update here: x.com/qi9098/status/…

watching /goal run in Codex for 101 hours feels kind of absurd. just letting Codex iterate on visual parity for a card game UI, reading tests, checking screenshots, and polishing small details. very interesting to watch.

the Sci fi fantasies I had as a kid is starting to become real

@thsottiaux 👀

Codex is for cosy Sunday evenings. Show me your cosy creations.

@Javi @OpenAIDevs @Dimillian the work you guys are doing is honestly amazing. I’ve been using ChatGPT and Codex since the early days, and you can really see the amount of effort and iteration being put into every new update 👌

@Zickox @OpenAIDevs @Dimillian I remember when we added image support, such a huge unlock!!

watching Codex work from the ChatGPT iPhone app feels surreal. real-time logs, simulator control, screenshots as visual evidence of changes, all streamed back with surprising clarity and fluidity. Fantastic work by the @OpenAIDevs team on this one 👏 congrats @Dimillian

@LionelAccount @OpenAIDevs @Dimillian Yep

@Zickox @OpenAIDevs @Dimillian In android , the permissions are set to default and you cannot change it. But in iOS you can change it to full access?

Actually same vibes here. There is definitely something magical about just being able to ask Codex to do some real work from your phone and oversee it. Also next release will make the thread tool calls much prettier and closer to the desktop app.

@manabiSRS @OpenAIDevs @Dimillian that actually makes a lot of sense. especially for design feel, animations, transitions, and the small UX details that are hard to fully validate from static screenshots alone

@Zickox @OpenAIDevs @Dimillian I need to see motion and a stream is nicer than only video (or screenshots). I also still value user testing. not for every or even most iterations, but I am not shipping anything without trying it, and that's something I do incrementally for design feel

@manabiSRS @OpenAIDevs @Dimillian realtime simulator streaming would be incredibly cool 👀 but part of me also wonders how much manual control we’ll actually need once Codex can already operate the simulator, navigate flows, and validate UI changes for us. curious, what would be your main use case?

@Zickox @OpenAIDevs @Dimillian I mean with realtime view of simulator, and ability to interact with it manually (like browser)