Changelog

Pro mode now starts with GPT-5.6 Sol, OpenAI's flagship model for complex builds, detailed design work, and difficult debugging.

Your existing model choice is preserved if you selected one yourself. You can switch the Pro model at any time from the build-quality picker or Preferences > AI Model > Quality Model.

Cloud projects now have a File Explorer. Open the Files tab in the editor's left panel to browse everything the AI built for you - and edit any file yourself.

• Browse your whole project. Folders load as you expand them, so even huge projects open instantly.
• Edit with a real code editor. Click a file to open it in a tab with syntax highlighting. Changes save automatically to your app.
• Stays in sync with the AI. When the AI finishes making changes, the file tree and your open files refresh on their own.
• Full file management. Create, rename, delete, and download files and folders right from the tree - folders arrive as a single archive.
• Upload from your computer. Drop files or whole folders onto the tree (structure preserved), or use the upload button.
• Markdown preview. Open any .md file and flip between source and a rendered preview.
• Drag to move. Reorganize your project by dragging files and folders in the tree.

Your files live on your app's own server, so what you see in the explorer is exactly what your app runs. Big or binary files (images, archives) show a summary card with a download button instead of opening as text.

Building a site is half the job - people need to find it. Your Playcode AI can now research real search data while it works on your project:

• Keyword research. Ask "what keywords should my site target?" and get real Google search volumes, difficulty scores, and search intent - not guesses. It can also pull the full keyword list any competitor ranks for.
• Backlink intelligence. See who links to any site: total authority, top referring domains, and the exact linking pages. Great for finding sites that link to your competitors but not to you yet.
• Live Google results. Check who actually ranks for a search today - and whether your site shows up.

Just ask in chat: "find low-competition keywords for my bakery site", "who links to my competitor?", or "where do I rank for handmade candles?". The AI pulls the data, reads it, and turns it into a concrete plan - new pages, better titles, link targets.

Available on paid plans. Each lookup uses a small amount of credits (about the same as generating an image), and repeated identical lookups within a day are free.

Your Playcode AI can now hand heavy work to background agents - and stay available while they run.

• It keeps talking while they work. Ask for research, an audit, or a build slice "in the background" and the AI starts a task, tells you the id, and keeps the conversation going. When the task finishes, the result lands right back in your chat.
• A tasks panel in the editor. See every background task with a live status - queued, running, finished - how long it's been going, and a stop button while it's alive.
• Every task is a real chat. Open any task to watch the agent work message by message, exactly like your main chat. Type into it to steer a running agent mid-flight, or message a finished one to continue where it left off.
• Built-in quality loop. For bigger builds the AI can run build -> audit -> fix cycles: a coder agent implements, an auditor agent reviews against your requirements and returns a concrete verdict, and fixes get their own task - before the result comes back to you.

Long tasks no longer look like a hang: if an agent hits a limit, you get the partial result and can send it a follow-up to continue. Tasks even survive our own deploys - they resume from their chat history instead of failing.

Background agents got a big round of polish the same day they shipped:

• Agents now build in parallel. The AI can run several build agents at once - each with its own slice of the work - instead of queueing them one after another. Big builds finish sooner.
• You can see what each task runs on. Every task now shows its agent type and model tier right in the tasks panel, and read-only helpers (research, code exploration) always use the economy model. A task never uses a more expensive tier than your chat is set to - upgrading is always your call, never the AI's.
• Research you can trust. A research agent must fetch real web sources now; if it ever answers without them, the result is clearly flagged so you know it came from memory, not the web.
• Smoother task navigation. The tasks panel has a proper header and close button, and the back button in a task returns you to wherever you came from - the chat or the list. Failed and stopped tasks say what actually happened instead of a generic line.
• Calmer credit warnings. Running low shows a quiet amber note above the chat box (dismissible), with a heads-up at 50% and 15% remaining - including on the Free plan, which previously got no warning at all. Out of credits is now a one-click "Top up" (or "Upgrade" on Free), never a red alarm.
• Clearer controls. Stopping a task is a stop square now (the old ✕ looked like "close"), task cards keep their title readable on narrow screens, your chat list shows a small spinner on any chat the AI is still working in, and undo works on file changes inside a task's transcript too.

The chat composer got a full redesign:

• One quiet frame. Softer corners, a lower profile, and every control in calm ink - the send button is the only thing that lights up.
• Notices slide out from behind. Credit warnings no longer stack a banner on top of the chat box; they dock behind it as a slim strip, so the composer always stays in front.
• Compact attachments. Files and images you attach are now small chips inside the frame instead of large preview cards.
• Little touches. A frosted-glass scroll-to-bottom button, a clearer sidebar icon, and a bit more breathing room in the header - especially on phones.

Every chat now carries its own receipt. Open the chat picker in the editor (click the chat name above the conversation) and the bottom of the menu shows two quiet numbers for the current chat:

• Spent in this chat - the total AI credits this conversation has used so far.
• Context - how much of the model's working memory the chat is filling.

When a chat runs long, the context number turns amber with a tip: long chats reply slower and cost more, so starting a new chat for the next task keeps you fast and cheap.

Nothing ticks while the AI works - the receipt only appears when you ask for it.

You can now pick exactly how Playcode builds - and swap the model behind each tier - right from the Choose build quality menu next to the composer.

Three tiers, your call.

• Economy - fast and low-cost, great for simple edits, fixes, and basic pages.
• Pro - the default: detailed design, the cleanest results, and the fewest errors.
• Ultra - new. Our most capable model, for the hardest, highest-stakes builds.

Pick the model per tier. Each tier has a small model chip - tap it to choose from a curated shortlist, with a cost signal so the trade-off is clear. Set your preferred Pro model once and it sticks.

Tune the context, too. A new Context dial controls how much of the conversation the agent keeps before it summarizes - turn it up for long files and deep sessions, or down to keep runs lean.

Builds on GPT-5.6 Sol (available in the Pro tier) now reuse cached context between steps, so multi-step agent runs are noticeably faster and spend fewer credits.

Why it matters. A long agent turn re-sends a lot of the same context on every step. Sol now caches that context, so the repeated parts are much cheaper - more of your credits go toward new work instead of re-processing what the agent already read.

Your teammates' names and emails now appear instantly in the People view and the share dialog - and keep working even when your connection drops.

No more waiting or blanks. Names and avatars for everyone you share a workspace or project with are kept in sync locally, so the roster renders immediately instead of blanking while it fetches. Open the People page or a share dialog offline and you still see who's who.