OpenAI used its September 29 DevDay keynote to make one thing cheaper: giving a recurring job to software and getting on with your day. For a team deciding what to hand over, that is the news. Whether it multiplies the team or only adds a subscription depends on three things: what the handoff costs, what OpenAI has shown as against what it has claimed, and who is still standing at the check.
This is the companion to Value-First AI Daily’s live coverage of the keynote, written for a reader who did not watch. It follows the show’s Top 3 in order. Every figure carries its source, and OpenAI’s own claims are labeled as OpenAI’s. OpenAI’s product pages returned errors when this was researched, so the closest primary source is its developer community announcements post, with press supplying prices.
Number 1: a standing job, checked at the door
Dots are assistants that live inside ChatGPT and keep working after you close the window. You give one a standing responsibility. Each one runs on a cloud computer and browser of its own, uses the apps you have connected, and returns to you when something needs a yes. You can talk to a dot from ChatGPT, Slack or Teams. Beside dots, OpenAI announced ChatGPT Space, a shared team home for files and context that people and their dots both use, and Pages, a document format that people and dots edit together, with charts and small tools inside (OpenAI Developer Community post).
The limits are the part to read first. The first dot is included in the Pro and Business Premium plans, according to BGR, SQ Magazine and Engadget. Extra dots are for sale at a price no source states. Enterprise is a beta, texting a dot is a limited US beta, teams of dots are listed as coming, and the launch leaves out the European Economic Area, Switzerland and the UK. The demos were OpenAI’s own, and no independent results had appeared by the time of writing.
The hosts noticed two things while watching. In the demo videos the person hands off the job and leaves the frame, and nobody comes back to inspect what was returned. And nothing shown, in their judgment, was beyond what a determined team could have assembled the week before. Both point at the same fact. OpenAI’s addition is the packaging: a standing delegation with an approval step, folded into a plan you subscribe to, where it used to take a build.
That approval step is the whole human contribution to the loop, so it is the piece worth designing on purpose. Who sees the check-in? What do they compare it against? What do they do when the result looks plausible and is wrong? The cost of a dot is a subscription line plus the attention of whoever reads its check-ins, and only the first of those appears on a price list.
Number 2: cheaper thinking, with speed sold separately
OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra, its flagship, on agentic coding, computer use and professional work, for roughly a fifth of what Astra charges per token. Press lists the price as $2 per million tokens going in and $10 per million coming out (SQ Magazine, dev.to). GPT-6 Sol had the same list price when it arrived on September 22, so the 6.1 update is the change and the price is not.
OpenAI is the only source for the near-match. Press names DeepSWE v1.1 as the coding comparison, and no independent run of it turned up. Sources also disagree on the model’s name and on whether it can be used yet: OpenAI’s Ultrafast documentation lists a “GPT-5.6 Sol” in preview, and one press snippet says live where the developer community post says coming soon. A team budgeting against it should confirm on OpenAI’s model page first.
Ultrafast is a paid top-speed tier. OpenAI says it returns results up to 8 times faster in Codex and 6 times faster in the API, for 6 times the standard price. For GPT-6 Astra, press puts that at $60 per million input tokens and $300 per million output (the-decoder, dev.to). It is live for Astra and coming for Sol, and in the API it is limited to US data residency and global processing (OpenAI’s Ultrafast documentation). ChatGPT Pro 500 is the consumer-side version: $500 a month, with 25 times the Plus usage allowance and Ultrafast included (BGR).
The smallest announcement may serve a team that sorts incoming work best. The Decisions API points a specialized GPT-6 Luna at a question with a fixed set of answers, such as which queue, which category or which step comes next, and returns one in about 150 milliseconds, against about 1.6 seconds for a standard Luna call (the-decoder, dev.to). Its listed price is $0.10 per million input tokens and $0.50 per million output, from a single source, and it is a limited preview with a broader release said to be days away. Deciding which reply an incoming Interest receives first is a job where a fast, narrow, inexpensive answer beats a clever one.
Here is what that does to the arithmetic. A task that was not worth automating at last month’s token price may be worth it now. Speed is separate money, and the team decides whether a given job needs it.
Number 3: who carries the token bill
With Sign in with ChatGPT, a person can spend their own ChatGPT subscription inside another company’s app, so the app’s maker no longer pays for those tokens. Sixteen launch partners are named, Notion, Devin and Vercel among them, it is open to Plus and Pro users, and the usage still counts against that person’s plan limits (dev.to). Plugin extensions let partners add sidebar entries, panels and file viewers inside ChatGPT, on all plans, and sites can host plugins on Business plans and up, and each user keeps their own login. Adobe and Figma are launch partners (OpenAI Developer Community post).
The OpenAI Marketplace is the enterprise door. Eligible enterprise customers can buy partner products and count the spend against an existing OpenAI commitment. OpenAI’s count from the stage was more than 30 launch partners, and dev.to counts 32, naming Adobe, Figma, Harvey, CrowdStrike and Baseten. None of the confirmed partners is a CRM vendor. A host wondered aloud whether HubSpot’s agents might show up there. Nothing yet supports that, and it is the hosts’ speculation, not a report.
A host’s read after the keynote gives the marketplace its weight. Plenty of people are building tools and apps, and until now there has been nowhere for anyone to go and find them. Whether that becomes a place where builders are paid for their ideas depends on terms nobody has published.
Episode — Value-First AI Daily
Live: OpenAI DevDay 2026
Chris Carolan and Nico Lafakis, September 29. The keynote runs from the first minute to about the 54th of a show that lasts roughly an hour. The hosts’ read of it follows, then Nico Lafakis shows a skill he is building.
Open the episodeWhat stays with a person
OpenAI’s own close, as heard on the show’s recording and not found in print, drew a line between a new renaissance and a new industrial revolution. The announcements themselves show where that line will be drawn in practice: in whether a team treats delegation as a way to free people for the parts only they can do, or as a way to need fewer of them. Nothing in a price list settles that. The first choice multiplies what a team can create, and the second only lowers its cost.
The show ended on a thread that points the same way. Nico Lafakis showed a skill in progress that reads a slide deck he gave no visual instructions for and builds an interactive 3D walkthrough from it people click through instead of paging. He said he would post about it. It has not been released, so for now it is a demonstration of a person with one deck multiplying what that deck can do.
Worth passing on?
