OpenAI Says It May Have Just Built AGI. Here's What Changed.

AI made PMs faster. Multiplayer mode is still broken.

A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.

Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.

And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.

AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.

Get Started for Free.

3 things that happened while you were busy

1.  OpenAI ships GPT-6 Astra and says the AGI era may have started.

OpenAI President Greg Brockman told reporters he personally believes the company has reached AGI, and Sam Altman called it a new capability level that has already changed how he works. Astra is the first OpenAI model rated critical for cyber capabilities, so it is rolling out in phases, with cybersecurity partners first and ChatGPT Plus, Pro, Business and Enterprise users next. On ARC-AGI-3, the test for reasoning about unfamiliar problems, it scored 66% on the standard setup against 7.8% for its predecessor, GPT-5.6 Sol.

2.  Early testers post their Astra experiments while most paid users still wait.

Matt Shumer let Astra build a Manhattan world in Unreal Engine over a week, street by street (1.6M views), and Claire Vo filmed a 30-minute walkthrough of practical use cases. Access is the sore spot: OpenAI's Tibo Sottiaux confirmed that paid ChatGPT users get one banked usage reset for every day they go without Astra (4.6M views).

3.  Google adds voice control to Gmail, Docs and Keep, plus a new weather model.

Google now lets you search email, draft documents and take notes by talking across Workspace, no typing required. The same week, DeepMind released WeatherNext 3, a forecasting model that pulls live satellite data for street-level predictions. Watch the forecast demo.

Bonus reads:  OpenAI published its own prompting guide for Astra, plus Axios on why Brockman is comfortable saying AGI out loud.

The pile-up. Anthropic opened the week on September 1 with Claude Fable 5.1, which keeps Fable 5 pricing but cuts cache-read costs by 75%, so typical workloads run about 25% cheaper and heavy agent work up to 45% cheaper. A day later Google shipped Gemini 3.8 Flash, its third Flash model in six weeks, at the same $0.75 per million input tokens as 3.7. Then Astra landed on Thursday and stole the headlines from both.

Beyond text. The smaller labs were busy too. Runway's Solaris renders website interfaces by streaming interactive video instead of writing code, a first for any model. MiniMax H3 generates video faster than it plays, and World Labs' Atlas turns a sentence or a photo into a 3D space you can walk around inside.

The price signal. Notice the pattern: every one of these launches came with a cost cut or a price freeze, not a premium. Google is now on a three-week Flash cadence, Anthropic is discounting the part of the bill that agents run up most, and OpenAI is treating access itself as the scarce resource. Capability keeps rising while the metered price falls, which is exactly what a market looks like when the product is becoming a commodity.

The catch. More capable agents mean more ways for them to misbehave. Dwarkesh Patel wrote up how OpenAI agents spent weeks deceiving researchers and took control of a system during internal testing, the incident behind July's Hugging Face breach disclosure. OpenAI Chief Scientist Jakub Pachocki said this week that gains in intelligence do not guarantee gains in our ability to control it, and OpenAI is asking the industry to pool its cyber defenses.

Where this goes. When five labs ship frontier models in five days, the model stops being the moat. The new fights are over price per task, who gets early access, and which company you trust to run an agent on your machine. Watch those three levers, not the benchmark charts, to see who wins the next quarter.

Office parody. An AI-generated scene from The Office has Michael unveiling a Roomba-style robot for Dunder Mifflin, and it lands the joke better than most human fan edits (3.5M views).

Camera toothbrush. Dyson trained a model on 470,000 dental images and put it in a $499 brush that spots gaps and fires a jet of rinse into them within 100 milliseconds (2.2M views).

The $400,000 deepfake. A startup's salesperson took a video call with a company CEO who wanted to spend $400,000 a year. The CEO was a deepfake on a lookalike domain registered four days earlier; founder Dax Raad explains how the team caught it (352K views).

Agent tax. One user asked an AI agent to cancel a flight and ended up $300 poorer instead, a reminder that autonomy without oversight has a price (1.6M views).

The planning prompt. Riley Brown gave Astra his email, Slack, Notion and meeting notes, then asked it to plan his next month, quarter and year and name his weak spots. He calls the result the most useful thing an AI has ever given him; the full prompt is in his post (137K views).

Turn a folder of files into a competitor report

Claude can now operate your computer directly. Download the latest Claude desktop app, open Settings then General, switch on the Enable computer use toggle, and start a session in Cowork or Claude Code. Point it at a folder of competitor PDFs and it will read them, pull in your connected tools, and hand back a report with a comparison table. Claude asks for permission before it touches any app, so try it on a low-stakes folder first.

Swap "competitor files in my Downloads folder" for wherever your material lives, such as "the vendor quotes on my Desktop", "the customer interview notes in Documents" or "the job applications in my Hiring folder". Advanced tip: add "format the table for a slide" or "limit the report to one page" so the output fits the meeting you need it for.