GPT-6.2 Astra: Early SVG Tests, OpenAI’s GPT-ADA Models and This Week’s AI News

Every few days a new set of “leaked” AI demos does the rounds, and it is hard to tell what matters. This week WorldofAI covered the run-up to GPT-6.2 Astra, plus a long list of updates from Anthropic, Google, Tencent and others.

Here is our plain-English round-up. We have separated what companies have actually shipped from what is still rumour, and added a few practical next steps you can take today.

What is GPT-6.2 Astra, and what do the early tests show?

GPT-6.2 Astra is the expected follow-up to OpenAI’s current GPT-6 Astra model. It hasn’t been released, but people with early access have shared outputs, and it reportedly appeared briefly in the model selector in the Codex desktop app.

The headline claim comes from an OpenAI update about an internal model, the one linked to a reported result on the Navier-Stokes maths problem. On BenchCAD 2.0, a benchmark for engineering design work, the video says:

If those numbers hold up, the interesting part is efficiency as much as raw power. A cheaper setting that beats the previous model at full effort could mean lower costs for everyday work.

Why are people testing GPT-6.2 Astra with SVG images?

SVG stands for Scalable Vector Graphics, a way of drawing pictures with code instead of pixels. Asking a model to “draw” a detailed scene in SVG is a popular test because it checks coding, spatial reasoning and attention to detail all at once.

The early Astra outputs shown in the video include SVG recreations of famous video game scenes, a detailed animal cell with labelled parts, and the well-known test of a pelican riding a bicycle. One Minecraft-style scene reportedly ran to 7.5 million characters on a single line of code, and was then animated by the same model.

In side-by-side portrait tests, the video’s host preferred Astra’s output to Fable 5.5’s. That is one person’s opinion on a handful of examples, not a fair benchmark.

What are OpenAI’s GPT-ADA model candidates?

Two unknown OpenAI model IDs, rc-ada-or-vm and rc-ada-sh-vm, were spotted on a model tracker and disappeared within minutes. Nobody outside OpenAI knows what they are.

WorldofAI’s guesses: “rc” could mean release candidate, “vm” could point to a virtual machine for computer-use tasks, and “Ada” is the name of one of OpenAI’s earliest small models. The host doesn’t think these are GPT-6.2 Astra itself, noting that a previous mystery model, Aurora Alpha, turned out to be an image model.

How is OpenAI using Astra to design real hardware?

OpenAI shared examples of its models being used in mechanical design, electrical engineering and simulation work. In one case, two OpenAI employees reportedly used GPT-6 Astra to design a small prototype device for talking to Codex, covering the enclosure, circuit board, interface and software, in under a week as a side project.

OpenAI is also said to be working with Chip Ganassi Racing on turning 3D scans into editable CAD models, and with Mazda on AI-driven vehicle design simulations. The company is clear that these are early experiments, and that the prototype still needed several rounds of human review before it could be built.

Two smaller OpenAI updates also landed. ChatGPT’s finance feature now supports voice, so you can talk through spending and saving after connecting your accounts. The host says he wouldn’t share his own financial data, which is a fair, personal call. And Unity launched a plugin for Codex in OpenAI’s plugin directory, giving Codex game-development skills inside Unity.

What changed in Claude and Claude Code this week?

Anthropic made business plugins easier to connect, and published a long list of Claude Code improvements from October. Here are the ones most people will notice:

• Auto-connect for plugins (beta, Team and Enterprise): admins can set up tools like Figma, Notion, Linear, Dropbox and Supabase once, and staff are connected automatically when they sign in to Claude.
• ultrathink: type it to use the highest reasoning effort for a single prompt. You can also set different effort levels for individual sub-agents.
• Better /usage: shows what is using up your limits and suggests fixes.
• Smarter compaction: long chats can be summarised before the prompt cache expires, so going back to an older session uses fewer tokens.
• Background commands: long terminal commands can carry on running instead of being stopped when they time out.
• Monthly API credits for Max, Ultra and Team subscribers, for building with Claude’s API and Agent SDK.

There is also good news for game makers. Claude Code had been refusing some requests that simply mentioned an existing game. From version 2.1.317, you can describe mechanics by referring to a well-known game and Claude should build something original, while still declining to copy the game itself. Our guide on building an original game with an AI coding agent shows how to brief it.

What is Anthropic’s FlashProve, and why does it matter?

FlashProve is open-source research from Anthropic that uses cryptographic proofs to show a response really came from the model a provider says it is running. It tackles a real worry: if you pay a third party for a premium model, how do you know they are not quietly using a cheaper one?

According to the video, Anthropic tested it on Meta’s Llama 4 Maverick, a roughly 400-billion-parameter model, on a single Nvidia H100 GPU. It reportedly generated proofs at about 110 tokens per second, with each proof around 19 MB. An earlier system’s proofs were said to be about 93.6 GB and took nearly a day to check.

It is early research, not a feature you can switch on. Until tools like this are common, there are simpler checks you can run yourself. Our new guide explains how to check which AI model a provider is really serving you.

Tencent HY4 and Crouton Alpha: which new models are worth a look?

Tencent’s HY4 is a new model with vision, aimed at coding, front-end work, finance, research and office tasks. Tencent’s own testing puts it at 75.2 on the Apex Agents benchmark, slightly ahead of DeepSeek V4.1 Flash, though the two trade wins elsewhere. Reported API pricing is $0.83 per million input tokens and $2.50 per million output tokens, and Tencent plans to release the weights.

Tencent also says it removed offensive hacking data from training while keeping defensive security data. The video shows HY4 building a playable zombie shooter in Unreal Engine over several prompts, not in one go.

Crouton Alpha is a free stealth model on OpenRouter with a reported 1 million token context window that accepts text, images, audio and video. People suspect it comes from Meta because it blocks the same 21 countries as Meta’s API, shares Muse’s reasoning settings and makes the same mistake on a rare Japanese word. Its tokenizer doesn’t match Meta’s published one, though, so the link isn’t proven.

What else happened in AI this week?

• Anthropic is reportedly using Claude to help design its own AI chips, with a team of about 15 engineers. OpenAI is working on its own chip too.
• Fable 5.5 early access keeps impressing testers. One example is a 60-second animated short film that an agent produced in just under two hours. The host predicts it will beat GPT-6.2; that is his opinion.
• Grok Bot added shopping tools for finding deals, comparing products and tracking orders.
• terminal-mac, a free open-source project, runs a macOS environment inside a terminal so an AI agent like Claude Code can use its own Mac without touching your desktop.
• DeepSeek V5 may be trained mainly on Huawei Ascend 950 chips, reportedly aiming for 2027.
• Binance confirmed a major security incident on 29 October. Reports link it to a developer installing a tampered copy of the Homebrew tool, with billions of dollars in crypto reported stolen. It is a sharp reminder to install developer tools only from official sources.
• Elon Musk predicts electricity, not chips, could slow AI by the end of 2026. It is a prediction, but power supply is a real constraint.
• Zyphra’s hierarchical attention (ZHA) projects up to 35 times faster attention at 10 million tokens. These are projections, not proven results.
• Robots: a snow-clearing robot, and Xiaomi’s CyberOne humanoid reportedly completing an eight-hour factory shift.

The video also discussed the US midterm results in Texas, a big data-centre state. The host’s view is that tighter rules could slow some projects, but that demand for AI computing isn’t going away.

What should you do about all this?

Don’t switch tools because of a leak. Use this week to tidy up how you already work. A short plan:

1. Wait for GPT-6.2 Astra to launch, then test it on two or three of your own real tasks before changing anything.
2. Check your Claude Code habits: run /usage, and keep the top effort level for hard problems. Our guide to routing simple tasks to cheaper models helps here.
3. Know what you are paying for. If you use a model through a reseller, run the checks in our model verification guide.
4. Give agents their own space. Run coding and computer-use agents in a separate environment, not on the laptop that holds your email and banking. Our guide on giving an AI agent its own computer walks through it.
5. Install tools only from official sources, and be wary when an AI agent suggests installing something unfamiliar.

Frequently asked questions

The bottom line

GPT-6.2 Astra looks like a real step forward, but until it launches the SVG demos and benchmark numbers are previews, not proof. The more useful news this week is quieter: better usage controls in Claude Code, early work on verifying which model you are really getting, and another reminder to keep AI agents in their own space. Your next step: pick one agent or AI tool you use this week, and check what it has access to on your computer.

Related posts

• Claude Mythos 5.5 Leak: Fable 5.5 Demos, GPT-6.2 Astra and This Week’s AI News
• Anthropic Recursive Self-Improvement: What the IPO Filing Change Means, Plus Fable 5.5 and This Week’s AI News
• How to Spot a Fake AI Model Leak Before You Download Anything

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio