Changelog

Apple shipped a major Expo Go update this week, which broke the phone preview for mobile projects: scanning the QR code showed an SDK-compatibility error on iPhone. Fixed on both sides - new mobile projects target the Expo version the App Store actually ships, and the AI now checks your project against the current Expo Go before handing you a QR code, upgrading the project first if it is out of date.

The QR code itself also got more reliable: it now always carries the secure exps:// link, so it opens in Expo Go on the first scan.

GPT-6 Astra - OpenAI's most capable model, built for the hardest end-to-end work - is now available on the Ultra tier, next to Claude Fable 5.1. OpenAI launched it yesterday behind a limited-access program; our access opened this morning and it went through our full verification the same hour: live tool-calling, prompt caching, and token accounting all confirmed on the real wire before it reached the picker.

Astra is priced identically to Claude Fable 5.1 ($10 / $50 per million tokens at the provider), and because its tokenizer encodes the same work to roughly a third fewer tokens, the same build tends to cost fewer credits. Claude Fable 5.1 remains the Ultra default.

Background agents (the tasks your AI delegates work to) can now look at images, capture page screenshots, and understand videos, just like the main agent. Visual work delegates reliably: captioning a photo gallery, writing honest alt text, matching a reference design, picking a video poster frame, or reviewing a page's layout no longer fails partway because the helper agent could not see what it was working with.

Available on vision-capable models across all task roles - explorer, coder, auditor, and researcher.

Gemini 3.8 Flash is now the Gemini option on the Pro quality tier. It launched today at the same price as its predecessors, and Google's own launch numbers put it ahead of 3.7 Flash on every published axis - software engineering, terminal coding, and multi-step reasoning. In our own SVG drawing benchmark it is measurably the strongest Gemini Flash yet.

It also costs you fewer credits than 3.6 Flash did. Google discounted this rate card after we first added 3.6 Flash, and our pricing had not caught up, so 3.6 Flash was charging roughly double what it should have. That is corrected: every Gemini Flash model now bills at the real rate.

If you had 3.6 Flash selected on a chat, it keeps working and stays on that chat - nothing you pinned has moved. It is simply no longer offered in the picker for new selections.

The cheap-and-fast tier is unchanged. Gemini 3 Flash is still the default there, and still the cheapest way to build.

Grok 4.6 is now the Grok option on the Pro quality tier, in place of Grok 4.5. xAI prices the two identically and calls 4.6 "the most intelligent and fastest model we've built" - and in our own SVG drawing benchmark the difference is visible: where 4.5 famously spent 668 tokens barely attempting the task, 4.6 actually engages with it.

If you had Grok 4.5 selected on a chat, it keeps working and stays on that chat - nothing you pinned has moved. It is simply no longer offered in the picker for new selections.

GLM 5.3 is now the GLM option on the cheap-and-fast tier, in place of GLM 5.2. Z.ai prices the two identically, and 5.3 is the newer and stronger model - Z.ai's own measurements show it doing markedly better agentic coding work while spending *fewer* output tokens than 5.2 at every reasoning-effort level, which on a reasoning model means the same work tends to cost you less.

It also brings a real reasoning-effort dial (low / high / max), which the Intelligence setting now drives.

If you had GLM 5.2 selected on a chat, it keeps working and stays on that chat - nothing you pinned has moved. It is simply no longer offered in the picker for new selections.

GPT-5.6 Luna is now available on the cheap-and-fast tier. OpenAI cut its price fivefold today - to $0.20 in / $1.20 out per million tokens - which makes Luna the cheapest model in the picker, below even Gemini 3 Flash. In our SVG drawing benchmark it lands right beside Gemini 3 Flash in capability, with a reasoning-effort dial that responds cleanly to the Intelligence setting.

We also passed today's OpenAI price cuts for GPT-5.6 Sol and GPT-5.6 Terra straight through - the same work now costs you fewer credits on all three.

Gemini 3 Flash remains the tier default.

You can now connect email services like MailerLite, SendGrid, Mailchimp, Klaviyo, and Postmark to a domain managed by Playcode.

These services verify your domain with DKIM records: CNAME records whose target contains an underscore, such as litesrv._domainkey.mlsend.com. Playcode's DNS editor rejected those targets as invalid, which blocked email authentication. Underscores are now accepted in CNAME, ALIAS, NS, and MX targets, so you can add the records exactly as your email service shows them.

The app now opens in your browser's language on your first visit - all ten languages, including the sign-in page. On the website, the language switcher now also lives in the top menu (next to Pricing on desktop, inside the menu on phones), and if the site isn't in your language, a small note offers the switch in your own language.

Also fixed on phones: on some Android screens the page could grow slightly wider than the display, which pushed the menu button half off the edge. The menu stays put now, and a new account button next to it takes you straight to sign-in - or to your projects if you're already signed in.

Claude Fable 5.1 is now the default model on the Ultra tier, replacing Claude Fable 5. It is Anthropic's most capable model, tuned for the hardest reasoning and long agent sessions - and it comes with a real cost improvement: reading cached context is 4x cheaper than on Fable 5. In practice that is most of what a long coding session pays for, so the deeper into a project you are, the more you save.

Token prices are unchanged. Claude Fable 5 remains available in the model picker if you have it selected.