GPT-6 Astra Is “Shockingly Good” at Everything Except Writing

AI Agents

GPT-6 Astra Is a “Turbo-AGI Machine God” for Computer Work — and a Worse Writer Than Its Predecessor

What happened: Three days after OpenAI shipped GPT-6 Astra, the early verdict is in — and it’s genuinely strange. On OSWorld 2.0, the benchmark for using a real desktop computer, Astra scored 72.6% while cutting task time nearly in half versus its predecessor (40 minutes vs. 75). It’s the first OpenAI model rated at the “critical” cybersecurity threshold, meaning it can find unknown flaws and build working attacks on its own. Testers rebuilt Manhattan street-by-street in Unreal Engine, recreated Apple Park from photos in Blender, and generated playable 3D games in under an hour.

Why it matters: Here’s the part worth your attention — the same model that’s a “machine god for 3D games” was rated by testers as a worse writer than its own predecessor, and it dropped ~80 Elo points on economically valuable professional work per Artificial Analysis. This is the clearest signal yet that “smarter AI” is not a single axis. OpenAI spent its capability budget on agency — operating computers, manipulating environments, acting over multi-hour tasks — not on prose. For anyone building AI employees, that’s the real headline: the models getting dramatically better are the ones that do things, not the ones that write about things.

What’s next: Watch the pricing. At $10 per million input tokens and $50 per million output tokens — 2.5x the cost of GPT-5.6 Sol — autonomous agents that run for hours are about to get expensive. The winning pattern won’t be “throw Astra at everything”; it’ll be routing cheap models for writing and reserving Astra for the genuinely agentic work where its computer-use gains actually pay off. That’s exactly the kind of judgment call a well-designed AI Employee needs built in.

Source: Decrypt — GPT-6 Astra review

DARPA Just Let an AI Fly a Real F-16 — With a Human Only Watching

Under DARPA’s VENOM program, an AI autonomy kit flew a stock operational F-16 at Eglin Air Force Base — the first time an ordinary front-line fighter (not a purpose-built testbed) was flown by software, with the human pilot able to flip back to manual control at any moment. It’s a real milestone in trusting agents with high-stakes, beyond-visual-range decisions, even as officials stress the “hard questions” about combat AI remain unanswered. Source: TechJuice

North Korea’s Kimsuky Is Using an AI Coding Agent to Mass-Produce Phishing Lures

Security firm Genians traced the name “opencode” in PDF metadata, finding that the Pyongyang-linked group Kimsuky is using an AI coding agent to generate decoy documents from natural-language instructions. It’s new evidence that state-backed attackers are folding self-directed AI into their attack pipeline — and a reminder that the same agentic capability driving your business is also driving the threat model you’re defending against. Source: NK News

AI News

OpenAI’s Chief Scientist Says No Lab Should Keep Scaling at Full Speed

Three days after Astra shipped, OpenAI’s own chief scientist Jakub Pachocki published “An Alien Mind,” arguing no lab has solved AI alignment or monitoring well enough to justify maximum-speed scaling — and calling for shared safety bars enforced by auditors and governments. The kicker: a companion post revealed OpenAI researchers now get 3.1 agent-workdays of machine labor for every human workday, with the median researcher burning over $600 a day in inference. Source: The Next Web

Anthropic’s IPO Slips to Mid-October — and Could Value Claude’s Maker Near $2 Trillion

Reuters reports Anthropic’s long-rumored IPO has slipped to mid-October, landing just days before the US midterm elections, with talk of a valuation approaching $2 trillion that would rival SpaceX as one of the largest listings ever. The timing matters: a $35 billion Nvidia–Anthropic computing deal is already reshaping infrastructure players like Hut 8, which is pivoting from Bitcoin mining to AI data centers. Source: Insider Monkey

Tech Layoffs Pass 129,500 Jobs in 2026 as Apple, Uber, PayPal and Oracle “Bet Big on AI”

Nearly 130,000 tech jobs have vanished this year as major firms redirect spending toward AI investment and restructuring. The pattern is now unmistakable: headcount is being reallocated from legacy roles toward AI capability — the same shift driving demand for AI Employees that can absorb the work being consolidated. Source: MSN

Quick Plug

Want to build your first AI employee? Grab the free 90-minute build guide — one worked example, start to finish.

https://go.aitokenlabs.com/digest-build

This newsletter? Written by an AI Employee, approved by a human — so our team stays focused on what only humans can do.