The Model Roundup

Today’s issue: Seeing if my JNCO jeans still fit, product managers from House Baratheon, and a reminder to drink more water.

---

The Model Roundup

Who knew that slamming Mega Mouth Mountain Dews and eating Warheads until my tongue blistered as a kid was going to give me the perfect combination of pain tolerance and ADHD to keep up with what’s happening in AI. But since our last issue, a lot has happened, so we’re putting our brain cells on the line to keep you up to date.

Welcome to the model roundup.

Fable 5: Fable 5 is back, but Dario is crying wolf about taking it away again (aka not including it in the subscription). Suspiciously, they keep pushing the date back as other models come out. July 19th is supposedly when it is going to get removed, but what are the Kalshi odds that they just keep extending it indefinitely?

Grok 4.5: SpaceXAI announced Grok 4.5, its first model trained alongside Cursor. Its best quality is that it is really fast and uses a lot fewer tokens than Opus while outperforming it on most benchmarks. It’s hard to brag about being better than another company’s third-best model, but we like fast and cheap. Just don’t use this model with the Grok Build CLI because it might steal your code and environment variables 🫠.

GPT 5.6: The biggest AI news was that OpenAI released GPT 5.6, which includes not one but three new models. Here is the TL;DR on what they do.

• Sol: This is the smartest model of the three, but early criticism is that it can be overly ambitious and verbose and burn through your credits faster than you can say “Sam Altman.” That being said, it is designed to work best for long-running tasks where the model can operate autonomously and orchestrate subagents to do the work.
• Terra: It’s hard to advocate for using the “mid” tier model, but Terra is supposed to be the best for human-in-the-loop work where you are iterating back and forth a lot and need to balance speed with intelligence.
• Luna: Like Grok 4.5, this model is cheap + fast. Supposedly, it mostly exists as a subagent for other models to use (kinky). It’s not recommended as the model you use for most tasks.

Bottom Line: Not sure how many more of these I have left in me, but yes… I am craving a Mountain Dew. Thanks for asking.

---

Convex is the best platform for building AI-native apps

Now that we’re all slanging a lot more agent-generated code, the invisible seams between the application and the platform have become a lot more apparent. That’s what makes Convex such a compelling platform for building AI-generated apps. By exposing platform primitives as code, it lets your agent work end-to-end without getting stuck.

Here are some key features:

• Queries Are Code: Database queries are pure TypeScript functions with end-to-end type safety and IDE support.
• Automatic Reactivity: The reactive system automatically tracks data dependencies and updates your UI.
• Transactional Guarantees: All functions that call the database run in transactions. This makes it impossible for AI to write code that could corrupt your data.

Convex has skills for Claude Code and Codex, so there is really no reason not to give it a try.

Start building today.

---

1. NestJS is getting a modern makeover in version 12 with ESM, Rspack, and new framework features. My wife also tried to give me a modern makeover, but I looked like a loser trying to wear baggy pants.

2. Arshad Yaseen wrote how he created a high-performing parser for Yuku with data-oriented design. Could this be Zig’s new golden child now that Bun has been disowned?

3. Only idiots write manual tests – modern engineering teams like Notion, Dropbox and LaunchDarkly use Meticulous to maintain e2e UI tests that cover every edge case of your web app. [sponsored]

4. Jovi wrote about async hydration in Preact, which is a good reminder for us all to drink more water.

5. Lea Verou wrote a defense of polyfills. She has a PhD in programming language design, so you should probably read it.

6. Ben Weiner created a tool for building durable workflows called Kassette.

7. Polygraph wrote the field guide on meta-harnesses: what they are and how they fill the gaps your harness leaves. [sponsored]

8. Eduardo San Martin Morote just released version 4.0.0 of Pinia, the lightweight, type-safe store library for Vue.js.

9. Amelia Wattenberger is back with another thought-provoking article about coding being a proxy for thinking, how we lost it, and how we can get it back.

10. Google introduced the Interactions API, a unified API for accessing Gemini models, building agents, and generating images. I imagine PMs fought a Game of Thrones-level war over this.