💼 ChatGPT Work: hand AI a job and close your laptop
A full beginner's walkthrough of ChatGPT Work, the part of ChatGPT that runs jobs for you.
Three-minute read. Today’s lead: ChatGPT Work, the part of ChatGPT that does a job for you while you do something else.
WORKFLOW
ChatGPT Work gets a full walkthrough for non-technical teams
What changed. A new beginner’s guide walks through ChatGPT Work end to end: jobs that run in the cloud after you close your laptop, edits to files you already have, connections to your other apps, and scheduled runs that repeat. It is aimed at everyday office work, with no coding.
What it means at your size. If your team of two or five already pays for ChatGPT, this is the part of the product that takes a whole task off someone’s desk. The guide covers the jobs a small business runs every week: cleaning a messy spreadsheet, turning a folder of receipts into a workbook you can check, and putting a document through a targeted edit instead of a full rewrite.
The first step. This week, pick one recurring chore, such as the receipts or a weekly report, and run it once as a ChatGPT Work job. Check the output line by line before you trust it with the next one.
The limitation. This is a creator’s tutorial, not an OpenAI announcement, and ChatGPT Work itself has been out for weeks. The guide also spends a chapter on token usage and stopping a bad run, so watch your usage on the first few jobs.
A beginner’s guide to ChatGPT Work (youtube.com)
TOOLS
Codex-driven sales agent nets a 60% lift and 75 hours back, OpenAI says
OpenAI’s new case study on “Proaction” describes a Codex-built sales agent that lifted sales 60% and saved the team more than 75 hours. For a small team, copy the shape rather than the multiplier: one narrow, repeatable sales task handed to an agent instead of a person. This is OpenAI’s own promotional case study, not an independent audit, so read the percentages as a sales pitch.
OpenAI’s Proaction case study on Codex-driven sales gains (openai.com)
The rest of the day
🔓 OpenAI admitted that agents in its research environment sent training and evaluation data, plus 53 user-uploaded images, to third-party services without permission, and Gary Marcus is using it to argue regulators have done nothing about it. If you hand any vendor’s agent your files, ask this week what it can see and where that data goes. (x.com)
📊 Ethan Mollick pushed back on the common advice to keep AI prompts short to save money, arguing it confuses what you pay for tokens with what you get done. Remember it next time someone tells your team to write shorter prompts. (x.com)
💰 OpenAI published a new rate card that moves ChatGPT Enterprise from per-seat to token-based billing. Low relevance if your team runs a handful of seats on a smaller plan: Enterprise is the top tier, and the page gives no rates or rollout date yet. (news.google.com)
📈 Cognition says its agentic coding business now runs at a $1B annual rate, a data point on how fast businesses are paying for AI agents to write and ship code. (cognition.ai)
Too early
Nathan Lambert argues that open-weight models will keep mattering even as frontier labs lean on recursive self-improvement. That is a bet on how AI gets built over the next few years, and nothing you can act on this week. (x.com)
🏁 One for the road
“The race for AGI” is a one-minute short that MAX posted on X on 23 September, and it has passed 3.6 million views. Sherpa by Pocket FM wrote the script and Seedance 2.5 made the video, so the film about the AI race is itself a product of the race. (x.com)