Issue 724

29^th August 2025

Written by Dave Verwer

Comment

With just under two weeks until Apple announces release dates for the v26 SDKs and operating systems, I thought we were done with Xcode 26 news. However, last night saw Apple release what is likely to be the final beta, with a significant change to the models available in its “Coding Assistant”.

The assistant now defaults to GPT-5 and, more significantly, adds support for using a Claude Pro or Max account. Read the release notes for the full list of changes, but it’s the assistant I want to talk about today.

I gave it a few fairly simple tasks that required both finding the code for what I requested and making the change. In my testing, Claude performed better in Xcode’s assistant than GPT-5 did. Firstly, with every change, it updated all affected test cases. It also added test cases where coverage would have been impacted without me asking it to. Secondly, it allowed for much longer conversations with the assistant. When I was using GPT-5, it told me that I needed to start a new conversation several times after just a few messages. I also got several timeout errors when using GPT-5.

All that said, you should take my experience as anecdotal rather than anything fundamental. Both models are extremely capable at making code changes in Swift, and it’s possible GPT-5 was just having some server load issues.

It feels strange to criticise what would have been a huge step forward in coding assistance just one year ago, but here I go. 😬 Unfortunately, my main point from Issue 716 still stands. An “assistant” like the one in Xcode 26 is useful, but it simply can’t compete with the full “agent” approach that comes with Claude Code, Codex CLI, or Gemini CLI. It can’t run the build, the tests, or any number of other hugely important command-line tools that the agentic tools use to assist the models with making code changes and checking the results.

I won’t dwell on this point, as there’s nothing more I have to add to what I wrote a few weeks ago. I will say that it’s good to see major model changes like this so close to the final release of Xcode 26, especially when those changes mean we can use the latest models. The bigger question in my mind is where Apple will take it from here, and whether we will need to wait until Xcode 27 to see something more agent-like.

– Dave Verwer

News

Four Corners: the first Playdate game written in Swift

It’s been more than a year since Rauhul Varma wrote about writing Playdate games in Swift, but it’s great to see they’re now appearing in the Catalog! 👍

Tools

Automating GitHub Action Workflows For Swift

Did you write a Swift package and stop just short of setting up CI for it? If you did, then Sam Deane has just the tool for you. It’s a command-line tool that inspects your package manifest, applies any customisations you’d like via a config file you can add to your package’s repository, and generates GitHub CI workflows for you. Sounds good to me!

Code

SwiftUI WebView

Sarah Reichelt takes an in-depth look at the new SwiftUI web view component. Unfortunately, the official sample code demonstrating the component no longer compiles, so she takes on the task of providing a working sample and detailing some of the quirks of using it.

Building AI features using Foundation Models

Majid Jabrayilov dives into the tricky problem of randomness when dealing with language models and describes how you can get predictable answers out of the Foundation model. Once you’re done, check out his follow-up, which covers structured content in model responses.

Jobs

Founding Senior Mobile Engineer @ Neon – Lead mobile innovation at Neon, a fast-growing startup building a privacy-first app that lets users earn from their phone calls. Shape products with real-world impact in a creative, mission-driven environment. – On-site (United States in NY) with some remote work (within US timezones)

And finally...

Can we make it a little more green? Do you have a few hours? 🤔