OpenAI announced more than 20 products and features at DevDay 2026, held September 29 at Fort Mason in San Francisco in front of roughly 2,500 attendees. And the models were almost the boring part. The real payload was infrastructure: an Agents API in public beta, cloud-based Codex environments, a paid speed tier. And a distribution play that points developers straight at ChatGPT’s 1.2 billion weekly users, a figure OpenAI itself put on stage. That last number is the one that should change how you plan your next build. DevDay 2026 wasn’t a model launch dressed up as an event; it was a landlord announcing new tenancy terms. And if you ship software for a living, you were just offered a lease.
What Actually Shipped: The Short Version
The announcements split cleanly into two piles, and the split matters more than any single item. On the ChatGPT side, OpenAI launched Dots (always-on agents with connected apps and their own cloud computer), ChatGPT Space and Pages for shared work, plugin extensions. And an enterprise Marketplace with 32 launch partners, per the Developers Digest recap. On the developer side, GPT-6.1 Sol arrived as an upgrade to GPT-6 Sol, the Ultrafast speed tier debuted, the Agents API hit public beta, Codex gained cloud environments and a voice-capable CLI. And a Decisions API landed in limited preview (CNBC’s live coverage has the full sequence).
Read the two piles together and a thesis falls out. OpenAI spent the day wiring ChatGPT into an operating surface that other people’s software plugs into. And wiring developers into an execution layer OpenAI itself hosts. The event page and the developer community announcement frame it as humans and agents collaborating.
I’d frame it differently: one enterprise assembling both sides of the marketplace, then charging for the rails and the fast lane on top.
The Real Shift: OpenAI Is Selling Distribution Now
Here’s the part I keep thinking about.
Altman said plugin extensions mean developers can build things that are, in his words, “essentially entire applications that feel native to ChatGPT,” distributed through OpenAI, according to CNBC.
Pair that with Sign in with ChatGPT, which lets people carry their ChatGPT account into other products. And the 32-partner Marketplace, and you’re looking at a full stack: identity, storefront, install base. Reaching 1.2 billion weekly users without buying a single ad is a genuinely good deal on paper.
Now the other side of the ledger.
When your login, your discovery. And your distribution all run through one vendor, the vendor owns the relationship and you own the invoice. Every app platform in history has run this play: generous terms at launch, a revenue share and ranking algorithm once developers are dependent. OpenAI hasn’t announced anything like that yet, which is exactly why the smart move is to treat ChatGPT distribution as a channel, not a home base.
Build the plugin, sure.
Keep your own domain, your own account system, your own email list.
Dots are the demand side of the same bet. They’re persistent agents that work across connected apps in the background, learn your preferences over time. And run on their own cloud computer, per OpenAI’s community announcement and The Verge, which notes Dots are initially limited to paid ChatGPT Pro, Business Premium, and Enterprise subscribers. An always-on agent that “do nearly anything,” in OpenAI’s phrasing, with its own machine and your app connections, is the most aggressive version of this platform I can imagine: the interface stops being your website and starts being a conversation your customer has with someone else’s agent.
The Builder Stack: Agents, Codex Cloud, and Speed You Now Pay For
The developer announcements are where DevDay 2026 gets practical.
GPT-6.1 Sol improves coding and computer use. And OpenAI’s community announcement puts its standard token prices at one-fifth of Astra’s. That’s the number to re-cost your API bill against this month. Ultrafast is a new paid speed tier. And OpenAI reports up to 8× faster token generation in Codex and 6× in the API with it, plus a 45% reduction in API time to first token and over 30% faster tool calls and workflows. Treat those as vendor-reported figures, not measurements you can bank on. But the direction is clear and worth planning around.
The Agents API entering public beta is the structural change. It ships with hosted execution, memory, tools. And multi-agent support, which means the queue-and-retry plumbing most of us hand-rolled this year becomes a managed service. Computer use in the Agents API lets agents operate software through its UI, the same way you’d click through it. I’ve driven enough flaky UI automation to tell you that anything operating through a screen instead of an API will break in production eventually; use it for software that forces you to, not for everything.
Two quieter releases deserve more attention than they’ll get. The Decisions API, in limited preview, uses the Luna model to classify inputs, route requests, or choose an action from predefined answers, which is exactly the cheap classification layer most agent pipelines need at their front door. And Codex Cloud runs tasks while your laptop is closed, with access across devices and reusable development environments, while the CLI update adds two-way voice, an /agents view, and worktree workflows. Coding work is becoming a background job you dispatch from your phone.
My contrarian read on the speed tier: it’s the most economically interesting item of the day, more than Dots.
Once speed is a separate line item, latency stops being a property of the model and becomes a budget line you justify per workflow. Interactive agent products will pay it, because time to first token is their user experience. Batch jobs that run overnight shouldn’t.
And the operators who know which of their pipelines is which just got a durable cost edge over the ones who flip Ultrafast on globally.
What This Means If You Ship Solo
For one-person operations and small teams, this event is good news wrapped around a decision you shouldn’t dodge. The good news: model prices on the coding-and-computer-use path just got cheaper, the agent plumbing just got hosted. And a distribution channel with 1.2 billion weekly users just opened to plugin builders.
The decision: how much of your product lives on someone else’s platform. And what your exit looks like the day the terms change.
One gap nobody addressed: nothing in the DevDay materials I’ve seen covers deprecation timelines or migration paths between these model generations. So assume the job of tracking that is yours. Build a small table of what you run, on which model, at what monthly token cost. It takes an hour and it’s the difference between reacting to the next price letter and pricing it in advance.
Do this before the end of the week.
Pull your last full month of API usage, price your five heaviest workflows on GPT-6.1 Sol at one-fifth of Astra’s standard rate. And mark each one “interactive” or “batch” so you know who actually needs Ultrafast. Then decide whether your next product ships as a ChatGPT plugin, on your own domain, or both. If you build AI automation and want a second pair of eyes on that audit, that’s what I do every day. And the DevDay recap is the right place to start reading.
