Noël Vranckx • 29 Sept 2026 • 9 min read

OpenAI DevDay 2026: six lessons for supply chain people

Always-on agents with their own computers, a multiple-choice API and a cheaper near-flagship model. What OpenAI’s biggest DevDay changes for planners, buyers and logistics leads.

Illustration of a 1950s telephone switchboard at night, a panel of jacks with a dozen patch cords looping to the desk and one terracotta cord running up to a jack on its own, a headset on a hook beside an empty stool, a wire tray of cards, a wall clock at three and a window with a crescent moon

*OpenAI announced more than twenty things today. Only a handful change how a planner, a buyer or a logistics lead will work. Those few matter a lot.*

Who in your team is working at three in the morning?

Nobody, I hope. But if OpenAI has its way, from today the answer is “a dot”. That’s the name it gave the always-on agents it launched at DevDay 2026 in San Francisco this afternoon. A dot runs on OpenAI’s best model, has its own computer and browser in the cloud, and keeps working on the goals you set while you sleep.

It was the headline of a keynote with more than twenty announcements: a cheaper model, a paid speed tier, an API that only answers multiple-choice questions, shared pages that people and agents edit together, and a new privacy set-up for companies.

I went through all of it with one question. What changes for someone who plans, buys or moves goods for a living? Six things, I think. First the facts.

What OpenAI announced

DevDay is OpenAI’s yearly developer conference. The 2026 edition, on 29 September, was its biggest, and the company now claims 1.2 billion users. The parts that matter for us:

dots. Always-on agents on GPT-6 Astra. Each has its own cloud computer and browser, connects to more than 4,000 apps and answers in ChatGPT, Slack, Teams, by text or by phone. You set what it may do alone, what needs approval and what is blocked. The first dot comes with the Pro and Business Premium plans. “Specialised dots” with a company identity are in pilot, after internal tests in procurement, invoice processing and customer service.
GPT-6.1 Sol. An upgrade of GPT-6 Sol, itself a week old. OpenAI says it comes close to Astra, the flagship, on agentic work, computer use and professional tasks, at a fifth of Astra’s token price: $2 per million input tokens, $0.10 cached, $10 per million output.
Ultrafast. A premium speed tier, for Astra now and Sol shortly: up to six times faster in the API and eight times in Codex, at six times the price. Astra Ultrafast costs $60 per million input tokens and $300 per million output.
The Decisions API. Built on Luna, OpenAI’s smallest model. You define a question with a fixed set of answers, send text or an image as context, and get back one answer with a confidence score in about 150 milliseconds. Limited preview today, broad release “in the coming days”, no price yet.
Computer use in the Agents API. The Agents API, in public beta since 10 September, now lets a hosted agent operate software through its screen, next to sub-agents and tool search. AWS offers the same inside its cloud as Bedrock Managed Agents.
ChatGPT Space, Pages and team tasks. A shared workspace where colleagues, ChatGPT and dots work on the same pages. Pages can hold charts that refresh from connected tools. Team tasks run recurring work on a schedule or when something changes, such as a new email.
Private Intelligence. Zero data retention with “private safety processing”: content flagged for a safety check is encrypted into storage you control, and only a hardware-attested runtime can open it for automated review. A “Private Inference” preview on confidential computing is due this autumn.

There was more: Codex in the cloud, a voice-driven command line, security scanning, plugin extensions, a marketplace with 32 partners, a $500 a month Pro plan, and half the previous allowance on the $200 plan for new subscribers. Every speed and benchmark figure above is OpenAI’s own.

Six lessons for supply chain

Here’s the frame I use to read a day like this. Each announcement is a lesson about where AI is going, and each lesson lands on one piece of supply chain work.

Lesson	What OpenAI shipped	Where it lands
The agent gets a computer, not just a chat box	dots, computer use in the Agents API	Supplier and carrier portals with no API
Sorting is being split off from talking	Decisions API	Exception messages, shared mailboxes, MRP output
Near-flagship intelligence gets cheaper every week	GPT-6.1 Sol	Bulk reading of contracts, confirmations and spec sheets
Speed is now a line on the price list	Ultrafast	Only where a person is waiting
The document becomes the workplace	Space, Pages, team tasks	The weekly S&OP pack, project pages
The data objection is being answered, the region question is not	Private Intelligence, market limits	Your legal and IT conversations

The agent gets a computer. Much of our supply chain still runs on portals: carrier booking sites, supplier confirmation screens, customs systems. None has an API a planner can use. An agent that can log in, click and read moves the line of what is automatable from “tasks with an interface” to “tasks a temp could do with a login”.

Sorting is being split off from talking. The Decisions API is a chat model with the chat removed: a fixed list of answers, one pick, a confidence score. That is the shape of most supply chain automation: which queue, which action, which priority. Readers of my Jev posts will recognise the idea, and The New Stack calls it OpenAI’s answer to Jev.

Near-flagship intelligence gets cheaper every week. GPT-6 Sol arrived on 22 September at half the price of its predecessor, and GPT-6.1 Sol a week later, closer to the flagship at the same price. On OpenAI’s own test of questions about complex PDFs with tables and small print, it beats Claude Opus 5.5 at less than half the cost per task. Contracts, order confirmations and spec sheets are that kind of document.

Speed is now a line on the price list. Six times the speed for six times the price is a fair trade when someone is waiting: a driver at the gate, a customer on the phone, a picker holding a damaged carton. It is a waste for a nightly batch. Decide per process, not per company.

The document becomes the workplace. A Page that refreshes from connected data, and that a dot can update, is a living S&OP pack. The team task that “posts a weekly project update” is the status report nobody enjoys writing. Keep a person’s name on every decision in it.

The data objection is being answered, the region question is not. Zero retention with private safety processing is a real answer to “who at OpenAI can see our data”. But dots are not available to Pro users in the EEA, Switzerland or the UK at launch, and Ultrafast has no EU processing endpoint. For a European company, that is the sentence to read first.

The first two lessons are the pair to build with. Here is what they look like on a plant floor.

Worked example: the shortage desk at Upshift’s Polish plant

Upshift is a fictional bicycle maker with two plants and five hubs. Every morning, the MRP run at its Bydgoszcz plant throws out about 400 shortage and reschedule messages (all numbers here are illustrative). A planner spends the morning sorting them: most are noise, some need a chase, a few need a decision. This is how I would rebuild that desk with what OpenAI shipped today.

Step 1: sort every message with the Decisions API

Each message goes in with its item, supplier and open order lines as context. The question has five answers: no action, chase the supplier for a date, propose a substitute or second source, reschedule production, or escalate to a person. Anything under 85% confidence goes to the person, whatever the answer.

Step 2: draft the chases with GPT-6.1 Sol

For the “chase” pile, Sol writes the message to the supplier: order, line, quantity, confirmed date, needed date, and the last confirmation PDF as evidence. The planner’s template sets the tone. Sol reads the hundred PDFs.

Step 3: read the portals with computer use

Three of Upshift’s frame and component suppliers only confirm dates on a web portal with no API. An agent in the Agents API, with the login held in a vault rather than in the prompt, opens each portal, reads the confirmed date and writes it into a Page. The screen recording is kept.

Step 4: keep the human where the money is

Reschedules and escalations stay with the planner. At 08:00 they open one Page: 40 items, each with the evidence, the confidence score and the time it was gathered. The other 360 messages are handled or parked, with a log.

Pile	Messages a day	Handled by	Tokens a day
No action	240	Decisions API	0.36 M in
Chase supplier	100	Decisions API, Sol drafts, portal check for the 30 portal-only lines	1.2 M in, 0.05 M out
Substitute or second source	20	Decisions API, then the planner	0.03 M in
Reschedule	25	Planner	0.04 M in
Escalate	15	Planner and buyer	0.02 M in

At Sol’s list price, the drafts and portal checks cost about $2.50 a day. The sorting runs on Luna, whose standard price would put 400 messages at about six cents; the Decisions API itself has no price yet. Call it under $3 a day in tokens for work that eats a planner’s morning.

Notice what that number says. The tokens are not the cost. The cost is the design (the five answers, the threshold), the test (a few hundred messages labelled by hand to check the confidence scores) and the supplier’s portal terms. That is where the planner’s expertise moves.

What it can’t do, and the traps

Europe waits. dots are not available to Pro users in the EEA, Switzerland or the UK at launch, while Business Premium gets them in all supported regions. Ultrafast supports US data residency and global processing only. Check before you plan a pilot.
A preview is not a product. The Decisions API has no published price, no stated maximum number of answers and no accuracy figures. Computer use in the Agents API is a day old. Build so the model behind it can be swapped.
A confident sorter with a broken input. In OpenAI’s own adversarial test, Luna gave its best guess instead of reporting a broken search tool in 28.7% of cases, against 2.8% for GPT-6.1 Sol. The Decisions API runs on Luna. Log the age and source of the evidence next to every decision.
Portals have terms of use. Many supplier and carrier portals forbid automated logins. Ask first, and give the agent its own identity rather than a planner’s login, which is also what OpenAI proposes for its specialised dots.
The vendor is moving fast, and not only forward. Prices changed twice in a week, the $200 plan lost half its allowance for new subscribers, and GPT-6.1 Astra, the flagship expected in October, was held back after it missed OpenAI’s own safety bar. Every figure in this post is OpenAI’s. Budget for change, and test on your own messages.

Count your exceptions before you buy a dot

This week, pick one person who sorts messages by hand every morning: a planner with MRP output, a buyer with a shared mailbox, a customer service lead with delivery complaints. Count one day’s messages and write down the five answers they actually choose between. That list is the first input for a Decisions API test, a Jev test or a plain prompt, whichever tool wins.

That is the real lesson of today. The models will keep getting cheaper, faster and more autonomous without any effort from us. Knowing which decisions in our operation are worth handing over, and which must stay with a person, is the part only supply chain people can do.

Which queue would you hand to an agent that never logs off, and what is the one action it must never take alone?

Sources

Illustration of a scale model of a small company on a workbench: two factories and a warehouse linked by roads with miniature lorries, beside a pair of callipers, a pencil and an open notebook

Read it, then try it

The posts on this blog are meant to be used, not only read. Each one ends in something to do: a prompt to run, a calculation to rebuild or a first step for this week. You learn what AI can do in supply chain by putting it to work on a real problem.

Meet Upshift

Most examples use Upshift, a fictional bicycle maker. It builds road and gravel bikes, each as a regular and an e-bike version, in two plants, and serves dealers and retail chains through five distribution hubs. Upshift isn’t a real company, and none of its data comes from one. Meet the company.

The dataset and the exercises

Behind Upshift sits a complete synthetic company dataset: products, suppliers, dealers, orders, stock, production and the transport network, all consistent with each other. Its dates follow a date you choose, so the data always looks current. The exercises on this site use the same company, each with a task, the files you need and a way to check your result.

A post keeps its example small enough to paste into a chat. The dataset and the exercises let you do the same work at the scale of a real company.

The dataset and the exercises are free with an account. Accounts are by invitation for now: join the waiting list from the sign-in screen and I’ll send you one.

Open the dataset Browse the exercises