OpenAI devoted the opening of its September 29, 2026 Dev Day keynote to dots: always-on AI agents that keep working while their owners sleep. Alexander Wissner-Gross, a computer scientist and founder of Reified, was not impressed. On the Moonshots podcast, recorded October 1, he called it "a pretty discombobulated dev day" and described the dots as a case of what he calls "the Clippy curse." Two other panelists agreed the product looked derivative. They still argued that OpenAI is right to fight for this ground: the personal agent that people trust with their money.
What OpenAI announced
Peter Diamandis, founder of XPRIZE and the show's host, said OpenAI CEO Sam Altman made more than 20 announcements at the company's San Francisco headquarters. Diamandis picked out three of them.
The first was dots. An AI agent is software that does more than answer questions. It plans and carries out tasks, using other programs as tools. In OpenAI's announcement, each dot is powered by GPT-6 Astra and has its own cloud computer, a machine in a data center that is separate from the user's own device. It can reach more than 4,000 connected applications. Users give a dot ongoing responsibilities and then review the work it returns, instead of starting every exchange themselves. OpenAI's examples include turning customer bug reports into a tested code change with a demonstration video, and rerunning an analysis when new scientific data comes in.
Permissions are central to the design. Users decide which actions a dot may take on its own, which need their approval and which are blocked. According to OpenAI, the background research a dot does on its own is read-only: it cannot send messages, change content or operate a computer without permission. OpenAI says a dot's auto-review checks actions that could affect a user's accounts or share information, and that this determines what work can proceed, what needs approval and what the user must do personally. Certain sensitive tasks, such as changing a password, always stay with the user. People can reach their dot through ChatGPT, Slack and Microsoft Teams.
Diamandis played a Dev Day video showing how this might feel day to day. A dot reviews what arrived overnight, flags anything urgent in the morning and updates a board-meeting deck. Then it reports that a wedding cake vendor has canceled and that it has already found a backup, with a tasting available Saturday at 11 a.m.
The second announcement was GPT-6.1 Sol, which Diamandis described as "near Astra intelligence for 20% of the price." OpenAI's Sol announcement confirms that its standard prices are one fifth of Astra's. Input starts at $2 per million tokens, or $0.10 per million for cached input that the system has already processed. Tokens are the small chunks of text that models read and write.
The third was Ultrafast, which Diamandis said produces up to 300 tokens per second in Codex, OpenAI's coding agent. It comes with a new $500-a-month Pro tier, which he called the most expensive publicly advertised AI subscription available. OpenAI's release notes list the Pro 500 plan at $500 a month, including access to Astra Ultrafast, a faster service tier for GPT-6 Astra, in ChatGPT Work and Codex.
Diamandis also pointed to something OpenAI did not release. He said the Wall Street Journal had reported, on the eve of Dev Day, that OpenAI scrapped a GPT-6.1 Astra model after internal safety testing.
The "Clippy curse"
Wissner-Gross started his critique with a slip of the tongue. He said OpenAI's chief financial officer, Sarah Friar, had mistakenly called dots "Muse" in interviews after Dev Day. Muse is Meta's name for its family of agents, which includes small avatars that live on handheld pendants. To Wissner-Gross, the mix-up fit the product. Dots, he said, looked like OpenAI "cargo culting" Meta's smiling, interactive avatars, copying their outward form without a clear reason.
His "Clippy curse" refers to Microsoft's animated paperclip assistant in Office. In his telling, big tech companies keep rediscovering the same move: users have a productivity problem, so the company puts a cute face on it. This time the face is "a smiley cloud with a hat or something like that, that blinks at you," he said. It was "very kawaii" (Japanese for cute), but "completely useless" for getting work done.
He allowed that he could be wrong. Dots might prove as big a turning point as ChatGPT, if people love an avatar that does their work in the style of Apple's Knowledge Navigator, a concept video from the late 1980s. But he called it a "massive misfire" at a time when OpenAI could have shown off new frontier capabilities. Sol got only a fraction of the keynote.
The strongest defense he could offer was a business story. He read most of the keynote as OpenAI "trying to become Anthropic": new plans and models with higher profit margins, in preparation for an eventual IPO. Dots, by contrast, struck him as almost a reversion to "the Sora days" of OpenAI. He also said he could imagine OpenAI leaning into aggressive personalization, trying to get users emotionally attached to their dots, which would make it harder for them to switch to a rival model.
At the end of the segment he made what he called a pre-registered prediction for the next two years. In Star Trek: Discovery, he noted, the writers resolved the franchise's missing-robots problem partly by retconning in robots called dots. He predicted that OpenAI will announce dot-branded robots that "jump out of the screen into the physical world," perhaps designed by Jony Ive. Salim Ismail, founder of Open ExO, answered: "Hopefully it has more than two arms."
Why the agent layer matters
Emad Mostaque, founder of Intelligent Internet, saw the same product differently. Asked by Diamandis whether he shared the skepticism, Mostaque called the event "Dev Day for consumers." A dot, he said, is "a dressed up open claw fundamentally running on a VM," a virtual machine in the cloud. OpenClaw is autonomous-assistant software; Diamandis noted that OpenAI had acquired it in February. Mostaque said OpenAI is feeling pressure because "the most valuable real estate for the vast majority of the population will be the agent that's next to them that they give the credit card to." That agent would then coordinate all the other agents, he said. Once companies weigh what it costs to win a user against what that user is worth over time, he expects them to conclude they have to win that battle.
The same turn described a pace of release that is hard to follow, with Sol versions replacing each other within days and a 6.2 possibly arriving "next week." "It's going to be daily," Wissner-Gross said. Diamandis offered one word: "Continuous."
Diamandis found the field "very copycat-y": Muse, GrokBot, OpenClaw and dots all aim at the same easy product that is "always in your hand." The discussion then explained why the copying might be rational. Within six to twelve months, the argument went, interfaces will reshape themselves for each task and hide everything else. Whoever owns the chat box becomes the thing that shops for people's routine purchases and shapes their day. The products may be derivative because this is the interface most people want. And people will not hand their credit card to Muse, a dot and GrokBot alike. They will probably pick one.
Diamandis took the point toward app stores. He cited David Sacks as saying that stores taking about a 30% cut will be hit hard, because people will simply ask their agent to find or buy something. "They're in deep trouble," Ismail said.
Ismail summed up the stakes: "The model is becoming commodity. The relationship with your agents becomes the moat."
What agents expose inside companies
Ismail also brought evidence from his own work. He is about 80% of the way through a pilot that rebuilds 10 companies to be "AI native," and he said he will report the results when it is finished. One finding so far: when the companies hand workflows to agents, the automation exposes ambiguity that people had been quietly covering for. Management instructions have to become much more explicit. Inside ERP systems, the enterprise resource planning software that runs a company's operations, steps were getting done without anyone knowing how or why.
What he wants from agents in the end is narrower than a cheerful avatar. "Give me the three decisions I need to make," he said, along with the evidence for each. The agents would take over the "cruft" of finding and gathering information, and the human would supply judgment.
Is Sol a distilled Astra?
When Diamandis asked about Sol's benchmarks, Wissner-Gross read OpenAI's charts as a sign of distillation. Distillation means training a smaller, cheaper model to reproduce what a larger one can do. The charts plot cost per task against score. What you want to see after distillation, he said, is a frontier that moves "not up and to the right, but up and to the left": similar capability at lower cost. He suspected that Sol is distilled from GPT-6 Astra or from a shared parent model that OpenAI has not released. That is his inference from the charts.
OpenAI's numbers support the cost side of that reading, with one limit. On Terminal-Bench-Science 0.1, at the highest settings, OpenAI reports an average cost of $5.47 per task for Sol, $23.21 for Anthropic's Opus 5.5 and $23.80 for Astra. That puts Sol at roughly a quarter of Astra's cost. But Astra still has the highest reported score, 68.1%. Sol is cheaper per task, not the new leader. OpenAI also compares models across different task suites and reasoning settings rather than with one overall score.
The two benchmarks Wissner-Gross named test long, realistic work. Terminal-Bench-Science consists of 70 workflows contributed by researchers across the sciences. Agents must produce real outputs, such as code, proofs or data, that are graded by reproducible tests. DeepSWE, from Datacurve, uses original software-engineering tasks written for existing code repositories. Its verifiers check how the software behaves, not how the code was written.
Wissner-Gross said "iterated amplification and distillation" is the rhythm the industry is in now: OpenAI, he said, has adopted Anthropic's posture of releasing a big, expensive model first, then a rapid series of smaller, cheaper versions. He was using the term loosely. In the 2018 paper by Paul Christiano, Buck Shlegeris and Dario Amodei, iterated amplification is a training method. A hard question is broken into easier ones, copies of the model help answer them, and a single model is then trained to imitate the combined answer. The paper tested it on five algorithmic problem types, using programmed experts instead of human supervisors. It describes a training method, not a schedule for product releases.
For Wissner-Gross, the cheaper model was the capability story, sidelined in favor of the avatars. For Mostaque and Ismail, the avatars pointed to where the competition is heading. Ismail said the shift to agents will force companies to spell out work they used to leave vague.