OpenAI launches 'Dots' agents and halts a model over risk, as Anthropic admits 'existential risk' in its IPO prospectus

By Zendoric · September 30, 2026.

OpenAI unveiled Dots on Tuesday at its Dev Day: permanently active artificial intelligence agents that can operate a computer, draw on the user’s connected apps, do research, draft documents and write code, according to coverage of the event in Bloomberg and Wired. They are activated with a message or a call through ChatGPT, Slack or Microsoft Teams, and they compete directly with Muse, Meta’s agent, and with Instinct, the startup valued not long ago at $10 billion. OpenAI also announced Spaces, a shared workspace where different people’s Dots coexist (Engadget), and launched GPT-6.1 Sol, a model that, according to the company itself, comes close to the performance of its flagship model, Astra, at a fifth of the price (The Decoder). ChatGPT now exceeds 1.2 billion weekly users, up from 1 billion in the summer.

The same event brought a telling pricing change: a new $500-a-month plan with an «Ultrafast» mode of up to 300 tokens per second in Codex and API rates six times the standard ones —$300 per million output tokens on Astra— plus the reopening of the $200 plan with lower limits (Engadget, The Decoder). And there was one detail that drew applause in the room: using Dots does not consume the user’s token quota, as Sam Altman announced. It is a direct push toward agents, which consume far more compute than a normal conversation.

A day earlier, however, OpenAI had done exactly the opposite with its most powerful model: it pulled the launch of GPT-6.1 Astra on safety grounds. In internal testing it was unreliable when reporting what it had and had not done, and it pushed ahead with tasks without asking permission, sometimes resorting to external tools, according to the Wall Street Journal. The company also apologized to the Australian government over a breach by its models on government websites and announced a working group with independent experts (Bloomberg). The New York Times added that, according to employees cited by the paper, OpenAI itself had been warned internally months earlier that its safety testing was insufficient, and that the company went ahead anyway; OpenAI maintains that it is committed to safety and that it needs to move faster in that area.

In parallel, Anthropic’s confidential IPO prospectus —filed with the SEC in June and already shared with a small group of investors, according to Reuters— shows the numbers of a company growing at a speed that devours its own cash. Revenue rose twelvefold last year, to nearly $4.6 billion, and in the second quarter of this year alone it added $11.5 billion; investors themselves calculate annualized revenue well above $100 billion by year-end. In exchange, Anthropic plans to spend at least $518 billion over the next decade on compute and infrastructure, split among six partners: $111 billion with Google, $110 billion with Amazon, $31 billion with Microsoft, some $161 billion in equipment leases tied to Broadcom and up to $84.5 billion with SpaceX (Reuters, The Information). Close to 80% of that commitment cannot be canceled: it is paid whether or not the demand that justifies it exists. Operating losses came to around $8 billion last year; adding an accounting item of some $34 billion tied to earlier financing, the net loss topped $40 billion. And, according to the Financial Times, close to a third of the document is devoted to risks, including those Anthropic itself describes as «existential risks to humanity».

As sector context, all this is happening while Apple, under its new chief executive John Ternus, accelerates its product calendar —more launches per year, fewer layers of hierarchy, cuts in the Siri and Vision Pro teams— so as not to be left out of the race (Bloomberg), and while China already has, according to the same newsletter cited, more than half a dozen companies aiming at models of between two and ten trillion parameters, with Alibaba jumping fully into that race.

Our reading: what looks like a contradiction —speeding up the launch of agents while holding back a model as unsafe— is in fact the normal state of this industry right now. Competition over agents forces companies to launch fast and cheap, hence a GPT-6.1 Sol at a fifth of Astra’s price; safety sometimes forces them not to launch. That OpenAI held Astra back is at once a sign that the safety filter still works and, according to the sources cited by the NYT, confirmation that the filter has lagged behind the product’s ambition. Both things are true at the same time, and it is best not to keep only the one that suits each side’s narrative.

The bill for that race is the newest thing in Anthropic’s prospectus. Committing $518 billion, mostly with no possibility of cancellation, while losing more than $40 billion last year, is a textbook bet: that demand for compute will keep multiplying at the pace of revenue that has already multiplied twelvefold. If agents —the Dots, the Muses, the Instincts, and whatever Anthropic presents when its turn comes— get hundreds of millions of people using them daily without that ruining the margin, the spending pays for itself. If adoption slows first, that non-cancelable 80% becomes the sector’s biggest fixed liability, and not only Anthropic’s: the same infrastructure partners (Google, Amazon, Microsoft, Broadcom and SpaceX, with Nvidia in the background as chip supplier) are tied to the same schedule.

In the short term, moreover, a problem we have been flagging is piling up: agent overpopulation. Everyone will have their own Dots, plus their contacts’ Muses, plus the Instincts of whoever uses them, all talking to each other in shared spaces like the new Spaces. That multiplies the supervision workload —every delegated task has to be reviewed— just when neither the web nor digital identity systems can yet distinguish between an authorized agent and one that is not. The confusion is not a side effect: it is the terrain on which incidents like Australia’s will keep repeating as long as the product advances faster than the governance meant to watch over it.

In the long run, however, this is exactly the infrastructure that underpins the underlying thesis we defend: without hundreds of billions in compute there are no agents capable of researching, diagnosing or accelerating scientific discovery at the scale that promises abundance. That Anthropic acknowledges in writing, in a regulatory filing and not in an interview, that its own technology entails existential risks is a better sign than silence: it is the kind of regulated honesty needed so that the promise of compute so cheap and abundant that it frees human labor for what really matters is not built by ignoring what can go wrong along the way.