Registro de mudancas

We fixed a caching issue that could keep the Preview showing old CSS after you changed a cloud project. Cloudflare was telling browsers to keep some Preview stylesheets for too long, so a normal reload could bring back stale styles instead of the latest version from your project.

New Preview requests now respect the project server's cache headers, so CSS changes refresh normally.

If your browser already cached the old response, clear it once: open the affected project or folder, open the browser developer console, go to the Network tab, turn on Disable cache, and reload the page. After that one reload, the stale Preview cache is cleared and future updates should work normally.

Playcode now gets to the usable editor faster. We split more of the heavy editor code into lazy-loaded pieces, load only the current language by default, and delay expensive startup work until the interface is already responsive.

Preview starts more gently. The browser no longer has to spin up the preview worker, bundler, and Service Worker immediately on page load. If you press Run, Preview still starts right away. If you just open a project and begin typing, the heavy work waits briefly so the page stays smooth.

Preview is also more reliable in embedded browsers. Playcode now checks for blocked browser features like Service Workers, cache storage, IndexedDB, cookies, and local storage. When a browser blocks something Preview truly needs, you get a clear message instead of a blank panel or a forever-loading iframe.

Sign-in and onboarding got a cleanup pass. The auth screens are cleaner, Google sign-in fallback behaves better when a browser blocks it, and new users get a smoother first-run flow before landing in the editor.

We also tuned the marketing homepage video so it waits for real user intent instead of competing with the first paint, which makes the first screen feel less stuck on slower devices.

You can now invite friends to Playcode and both of you get credits. Your friend receives 10 welcome credits the moment they sign up with your link, and you receive 40 credits when they subscribe.

Find your personal invite link in the project menu, right under your credits meter. You have 3 invite passes; every time an invited friend subscribes, one pass comes back, so your best invites never run out.

The invite dialog shows who joined and what you have earned, with friends' emails masked for privacy.

Usage History now shows where every AI credit goes. Click the info icon on any row for a full breakdown: input, output, and cache tokens, each with its own credit cost.

It answers the question we hear most: why did a short message still cost credits? The first message in a chat saves your whole prompt and tools to cache once (that is the big number), and every later message reuses it at a large discount. Now you can see exactly that.

The activity list also got a proper Load more button and a running count of how many records you are viewing, so you can browse your full history.

OpenAI's newest general-purpose model, GPT-5.5, is now available alongside 5.4 in Playcode's model picker. Pick it as your Quality model in Menu → Preferences → AI Model → Quality Model if you want to try it.

Heads up on cost. Per OpenAI's May 2026 rate card, GPT-5.5 is roughly 2× the per-token price of GPT-5.4 across input, cached input, and output. 5.5 shares the same tokenizer family as 5.4, so net effective per-request cost lands at about 2× what 5.4 charges for the same work. Real, not theoretical.

Not the default - Claude Sonnet 4.6 keeps that slot for now. We add new models as they ship so you can A/B them against your usual pick on real work; no model is unilaterally better than the others on every job, and the right answer depends on what you're building.

Honest take: no AI today gets everything right. Even Claude Opus, which still does UI design better than any model we've tested, takes its time and sometimes over-engineers a request. GPT-5.5 has its own trade-offs. Run the same prompt across two or three models when it matters; trust your own eyes on the output.

Earlier today we moved the default Quality model to Claude Opus 4.8. A few hours of real-world Pro usage made it clear that 4.8 burns credits roughly 2.6× faster than Claude Sonnet 4.6 for comparable output on typical website-building work.

Concretely: on the entry-level Pro plan, leaving 4.8 as the default would have torn through the monthly credit allowance in a handful of busy days. That's not the experience we want anyone to land on by default.

So we're walking it back: the default Quality model is now Claude Sonnet 4.6. Strong reasoning, sharp design, well-balanced per-request cost. Your Pro credit budget goes meaningfully further.

Opus 4.8 stays available for the work where the extra capability is worth the cost. Switch any time in Menu → Preferences → AI Model → Quality Model. Same place for Opus 4.6 if you want the previous default.

Honest about the back-and-forth: we set 4.8 as default thinking the quality jump would be worth the spend. The first hours of paid-user data said otherwise. Better to adjust quickly than to leave you on a setting that doesn't fit your plan.

We've moved the Quality model from Claude Opus 4.6 to Claude Opus 4.8 - Anthropic's newest Opus. Sharper reasoning, cleaner code, better designs. The difference in output quality is immediate and visible.

Honest about cost. 4.8 charges the same per token as 4.6, but its tokenizer packs roughly 25-48% more tokens into a typical request (the upper end on image-heavy work), so expect spend on Pro work to scale up by about that much. We made the call anyway: we want Playcode's default AI to be the best one available, not the cheapest.

Prefer lower cost? Switch the Quality Model yourself - it takes about 10 seconds:

• Menu → Preferences → AI Model → Quality Model
• Pick Claude Opus 4.6 (the previous default) - same family, lower per-request cost
• Or Claude Sonnet 4.6 - cheaper still, excellent for most tasks

Until today, the AI agent could fetch a webpage's text (article-style, cleaned via Markdown) but not its raw HTML. That was a real gap when you'd say "redesign this site" - the agent needed the actual DOM, classes, and layout, not a flattened text version.

Now there's a dedicated WebFetchHtml tool. Give the agent a URL, it downloads the page's raw HTML (up to 5 MB), saves it as a file inside your project under temp/, and reads it back chunk by chunk - the same way it reads any other file in your project.

Practical example: "Fetch https://stripe.com/pricing and redesign our pricing page in that style." The agent now actually sees Stripe's grid, their semantic landmarks, their spacing patterns - not just the text on the page.

Safety: same SSRF protections as the existing fetcher - localhost, private IPs, cloud metadata endpoints are all blocked. The HTML is never inlined into the conversation (that would blow the context budget); it's only ever read on demand from the project file.

The chat attach menu (the paperclip) now accepts zip and tar.gz archives, on top of images, videos, and PDFs. Drop one in chat, the agent gets the URL, and you can ask it to do something with the contents.

Where this really shines: Cloud projects. The agent has a real shell, so it can curl the archive into your workspace and unzip it directly. Practical uses:

• Import an existing site - drag its source zip in, ask the agent to extract it and use it as a starting point.
• Hand the agent an asset bundle - a folder of images, fonts, or design exports, packaged once, dropped in chat.
• Move work between projects - export from one, drop into another.

On Browser projects the agent sees the archive's URL but can't extract it in-browser (no shell). For full unzip-and-use behavior, drop archives into a Cloud project.

Up to 150 MB per file. The archive contents are never sent to the AI model directly - only the URL is referenced in the conversation, which keeps your costs predictable and avoids hitting model file-size limits.

Two new thinking models are now selectable in the chat model picker, talking to DeepSeek's official API directly:

• DeepSeek V4 Pro - frontier-level reasoning with a 1M-token context window. Strong at code, math, and long multi-step tasks. A serious daily driver. Text-only.
• DeepSeek V4 Flash - the same reasoning shape, tuned for speed and high-throughput work. 1M context, text-only.

Heads-up: these models do not see images. Pick a vision-capable model (Sonnet, Gemini, GPT-4o) if your task needs to read a screenshot or a PDF.