OpenAI says its best model, GPT-6 Astra, now scores 80.6% on a test of your life admin. Doctor appointments, bills, government forms, all the stuff that eats your Saturday.

The test is called LifeVal, and GPT-5.2, released last December, scored 49.3% on it. That's a huge jump, and I was ready to hand over my Saturdays.

Then I opened the PDF. OpenAI built LifeVal, graded it with its own GPT-5.6 and only tested its own models. 

That's like setting your own exam, checking your own paper and then announcing you topped the class.

Also, the survey that decided which chores to test was of Americans, in English, and free ChatGPT doesn't run the 80.6% model. Free users started getting GPT-6 Luna in October.

So I handed free ChatGPT two very Indian problems. Results after the news and partners.

If you read till here, you might find this interesting

Presented by Attio

We gave GPT-6 Astra the same job twice: analyze 100 sales calls and draft the follow-ups, once in Ask Attio and once in ChatGPT.

Ask Attio finished in 4 minutes 49 seconds. ChatGPT took 26 minutes 11 seconds, about 5.4 times as long.

The most ambitious GTM work requires more than frontier intelligence. It requires architecture that can put that intelligence to work at scale.

Learn about Attio’s hybrid architecture.

Presented by Hubspot

November 19th, 11am ET / 8am PT.

Pebl’s Chief Business Officer, Dutta Satadip, joins The GTM Garage to show how Pebl merged Operations and GTM under one leader, and how it aligned goals and incentives across the customer lifecycle.

You’ll also see how Pebl uses AI across that lifecycle and the stack behind prospecting, pipeline, and handoff to onboarding.

15 min conversation. 15 min open Q&A. For Seed to Series A teams.

Request Your Spot

AI is now hunting security holes for a living.

OpenAI’s pipeline checked 2,308 possible Linux kernel bugs and wrote working exploits with candidate fixes for 948, at about $1,280 a bug. Claude, meanwhile, got credited in Xen’s security advisory for two flaws that let one virtual machine read another’s memory.”

Almost nobody books an Uber through an AI agent yet. 

Uber says just four rides were booked by AI agents in its busiest October week, against about 320 million trips and orders. Four :)

Apple is rationing AI inside Apple.

Some engineers' daily AI budget dropped from $300 to $200, with $100 under discussion and some non-engineering teams possibly down to $25.

Grok can now phone a company and cancel a subscription for you, once you approve the call. → Read more

OpenAI is buying Pierre Computer, a Git-hosting startup whose diff and file-tree tools already run in Codex, in a direct shot at GitHub. → Read more

Anthropic signed its first content deal with a news publisher, giving Claude live AP election data. → Read more

Every newly married Indian gets to enjoy this one:

About 2,000 words came back, with a timeline working backwards from my honeymoon.

✅ Fees: It quoted July's new passport fees, ₹2,500 normal and ₹5,000 Tatkaal. Plenty of sites still show the old ₹1,500.

✅ The catch I didn't ask for: It remembered from an old chat that my passport expires in March 2027, and Indonesia wants six months' validity on arrival. That one catch alone could have saved the honeymoon.

❌ The order: It said passport first, in case Aadhaar rejects the change without proof. Later it listed the marriage certificate as proof UIDAI accepts, so all those weeks of waiting on a passport would have been for nothing.

✅ Tatkaal: It suggested Tatkaal if normal processing can't beat my travel date. I was sure that was wrong, since Passport Seva's not-eligible-for-Tatkaal list says "any change in name", so I asked for the official page.

It found the instruction booklet, with a row for a woman changing her surname after marriage and Tatkaal allowed, so only major name changes are blocked. Same website, two PDFs, opposite answers. Peak sarkari.

❌ It folded: It then apologised for having "incorrectly suggested" Tatkaal was off the table, which it never said. I was the one who got it wrong, and it still said sorry. Too sanskari.

✅ Don't pay: It flagged the fake domain and the midnight deadline, and pointed out the ₹1,000 fee is genuine, which is why the message sounds believable.

✅ My PAN: It didn't pretend to check. It sent me to the official e-filing portal to check the link status and told me to stop posting my PAN in chats.

❌ The ₹10,000: It called the fine misleading without saying where it comes from. Section 272B has a ₹10,000 penalty for PAN defaults like not quoting a valid PAN, so both scary numbers are borrowed from the law, and that's how it fools your uncle.

Pangram Image

Tells you if an image is AI-made, with a heatmap of the generated-looking parts. Pangram says it catches 99.2% of AI images, though in a demo shared by its CEO, it answered "unsure" on two images Engadget called obvious fakes.

Floot

Build an app by chatting in ChatGPT or Claude, now with Stripe payments. That startup idea you had in the shower can now take money.

LifeVal says 80.6%. My scorecard is 5 ticks out of 8, about 63%, which is better than I expected from free ChatGPT.

What still bugs me is the apology. I pushed back sounding sure, and it took the blame for a mistake it hadn't made.

LifeVal can't catch this, since it grades answers against a checklist written in advance. So for anything with a form or a deadline:

Has an AI, or a government website, ever given you two different answers for the same form? Hit reply.

Until next time,
Vaibhav 🤝