Composer 2.5 makes waves

Hi friend!​
​
​Google's big developer conference I/O happened last week and it was, as expected, a bit of an AI fever dream. There's been a ton of commentary on it already, so I'll spare you the details, let me just say - it feels like they have little focus right now and all of their releases are significantly behind the rest of the industry - see i.e. how Antigravity stacks up against Cursor or Codex. But then again - it's just Google being Google - as per Ben Thompson in Stratechery:

At least in my bubble, the show definitely got stolen by Cursor's new model release: Composer 2.5. I'll take a closer look at it in the notes section of today's newsletter.
​​
​Thanks for reading,
Robert

Composer 2.5 quickly became my go-to coding model

Choosing the right coding model usually means trading off capability, speed, and price. Cursor’s new Composer 2.5 makes that tradeoff feel much less of an issue.

It was trained on the same open-weights basis (Kimi K2.5) as its predecessor Composer 2, but with significant improvements to the post-training process, way more synthetic training data, and of course: more compute. This compute is accessible to Cursor through its partnership with (and soon likely acquisition by) SpaceXAI.

In Artificial Analysis' AI Coding Agent Benchmark for performance it took third place right behind GPT-5.5 and Opus 4.7 but does so at a ~60th of the cost ($0.07 vs. $4.14/$4.33).

I took it for a spin over the last days and did my own tests - how well does it work as a though partner when architecting a feature, how good is it at one-shot vibe-coding a 3D game, how does it work to use it to architect and build a new project from scratch.

Across the board, it performed incredibly strong - to be honest I could not notice a true capability gap compared to my previous daily driver GPT-5.5. And even if there was one, it was masked by the speed (especially when using the Fast version) that makes it much easier to stay in focus and iterate faster.
Plus - at the price point and with a Cursor subscription, the entire concept of token/price anxiety is gone with this model.

On X this sentiment was shared by many users - I know I'm in a bubble, but my timeline was filled with positive sentiment and even people saying they were switching back to Cursor simply for the model.

One small issue I've had with its personality is that it tends to be a bit "lazy" and not follow through being in the driver seat - it will i.e. make a change and then instead of running the dev server and testing it, tell you how to do it manually. A simple instruction line in your Agents.md file can fix this, however.

I know it's hard to overcome a habit, so if you're stuck with your flows inside of Claude Code or Codex and haven't tried Composer 2.5 yet - please do so, I promise it's worth it!

Headlines

• ChatGPT can now work directly with you in PowerPoint

ChatGPT @ChatGPTapp Have you ever thought to yourself: I really don't want to make this PowerPoint. Good news: ChatGPT can now create and edit presentations directly in PowerPoint. Build, update, understand, and polish presentations directly in PowerPoint while keeping slides editable. Now in 8:32 PM • May 21, 2026 310 Retweets 3798 Likes Read 198 replies

​

• ​Codex Appshots is a way to bring another apps context into Codex easily - part of a list of new features
• ​Warp now lets you coordinate agents across harnesses with support for Codex, Claude Code, and their own Oz

Miscellaneous

• ​Using AI to write better code more slowly - who says AI needs to make you faster, you can also just focus it to be better
• ​Shortcut Pro beats Claude & Copilot at Excel - no tool is perfect yet, but a niche tool beats the big players