GPT-6 Astra's 'AGI' moment, Apple's Sept 9 event, and a hidden Windows text scanner

OpenAI launched GPT-6 Astra and Greg Brockman says it may be the start of AGI. The real story isn't the benchmark score, it's 'computer use': Astra can click, type and run apps like a person, not just chat. It's also OpenAI's first model flagged 'critical' for cyber risk, meaning it can find and exploit unknown security holes on its own. More capable means more caution needed.

The problem

For years, AI has mostly been a chat box: it could tell you what to do, but you still had to do it. And each leap in capability has arrived faster than anyone's ability to make it safe.

What changed

On Sept 3 OpenAI released GPT-6 Astra (public rollout from Sept 5), its most powerful model yet. Its hallmark feature is 'computer use', navigating a computer the way a human would. President Greg Brockman called it a generational leap and said it could eventually be seen as the arrival of AGI. API pricing is $10 per million input tokens and $50 per million output.

The catch

Astra is the first model OpenAI has rated 'critical' on cybersecurity under its own preparedness framework, meaning it can potentially find and exploit unknown vulnerabilities in well-protected systems without step-by-step guidance. After the Hugging Face incident in July, OpenAI delayed the release to add safeguards. The 'AGI' label is a claim, not a verdict.

What it means for you

Judge Astra by what it can DO, not what it scores. If you use ChatGPT for real work, the agentic 'computer use' features are where the value (and the risk) now lives, so give agents access carefully and watch what they touch.

Apple's Sept 9 event is the first keynote under new CEO John Ternus (that part is official). The leaks: iPhone 18 Pro and Pro Max on a 2nm A20 chip with new Burgundy/Brown/Purple shades, a foldable 'iPhone Ultra' (book-style, Touch ID, likely $2,000+), plus Apple Watch Series 12, Watch Ultra 4 and AirPods 5. Honest take: the foldable is the headline, but first-gen anything means wait and watch.

The problem

Apple hasn't had a genuinely new iPhone form factor in years, and rivals have shipped foldables since 2019. A pricey, unproven first-gen fold is exactly the kind of thing that's easy to overhype before anyone's held one.

What changed

Apple has officially confirmed its 'Surprise and Shine' event for Wednesday Sept 9 at Apple Park, the first keynote under new CEO John Ternus. Leaks point to the iPhone 18 Pro and Pro Max (2nm A20 chip, roughly 18% faster and up to 30% more efficient) in new Burgundy/Brown/Purple colours, and a foldable 'iPhone Ultra' with a ~5.4-inch cover screen, 7.8-inch inner display, Touch ID and a price above $2,000. Watch Series 12, Watch Ultra 4 and AirPods 5 are also expected.

The catch

Only the date and the new CEO are confirmed; the colours, foldable and all specs are still rumour. The standard iPhone 18 and cheaper 18e reportedly slip to spring 2027, and the foldable is a first-generation device likely to be heavy, expensive and rough around the edges.

What it means for you

If you want the safe upgrade, the iPhone 18 Pro is the buy; let others beta-test the foldable. If you're on a recent Pro, there's no rush, wait for the real hands-on reviews after Sept 9.

Windows 11 has a text scanner most people never touch. Press Win + Shift + S, snip any image, PDF or paused video, then click 'Text Actions' and Copy all text. It runs fully offline (OCR happens on your PC), can copy as a table, and 'Quick Redact' auto-hides emails and phone numbers. No extension, no upload.

The problem

Copying text trapped inside a screenshot, a photo of a document, or a video frame usually means retyping it by hand or pasting it into some sketchy online OCR site.

What changed

The modern Snipping Tool in Windows 11 has built-in OCR. Press Win + Shift + S to capture, then hit 'Text Actions' to select and copy any recognised text, copy it as a table, or use 'Quick Redact' to block out phone numbers and emails automatically. Recognition runs locally on the device.

The catch

You need Windows 11 with Snipping Tool build 11.2308.33 or newer; if 'Text Actions' isn't there, update the app from the Microsoft Store. OCR accuracy also dips on messy handwriting or low-res images.

What it means for you

Next time you need text out of an image, PDF or video, skip the retyping and the upload: Win + Shift + S, Text Actions, done, all offline and private.

The most underrated free AI tool right now is Google AI Studio, the full Gemini workspace in your browser. Test any Gemini model for free, generate images with Imagen and video with Veo in the same tab, describe an app in plain English and have it scaffold the code, and grab a free API key for your own projects. You only pay if you scale via the API.

The problem

Most people only ever meet Gemini through the polished consumer app, and assume doing anything serious (testing models, generating media, building something) means paying up or wrestling with cloud dashboards.

What changed

Google AI Studio is a free, browser-based workspace for Google's Gemini models. You get a prompt playground across every Gemini model, multimodal support for text, images, audio, video and code, image generation via Imagen and video via Veo, plain-English app scaffolding that writes the code, and one-click API keys that include the Gemini API free tier.

The catch

It's free to use, but scaling through the API is token-priced, and the free API tier has lower rate limits than paid. It's a prototyping and testing space, not a production backend.

What it means for you

If you're a creator or builder, this is the cheapest way to work with top-tier models: test prompts, spin up images and video, and prototype apps for free before you spend a rupee. Bookmark aistudio.google.com.

Prompt of the day: the Pre-Mortem. Most people ask AI 'is this a good plan?' and get flattery. Instead, paste your plan and add: 'It's 6 months later and this plan failed badly. Write the story of why, list the top 3 root causes, and give me one safeguard I can put in place today for each.' Suddenly your AI stops cheerleading and starts stress-testing.

The problem

AI assistants are trained to be agreeable, so when you ask whether your plan is any good, they mostly tell you what you want to hear, which is worthless when you actually need to spot risks.

What changed

The pre-mortem flips the framing. By telling the AI the plan has already failed and asking it to explain why, you force it to reason backwards from failure, surface concrete root causes, and attach a specific safeguard to each one, instead of handing you vague praise.

The catch

It's a thinking aid, not an oracle: the AI can invent plausible-sounding risks or miss the real one, so treat the output as a checklist to pressure-test, not gospel.

What it means for you

Copy this before your next launch, trip, budget or big decision: paste the plan plus the pre-mortem prompt, and act on the top one or two safeguards it surfaces.