OpenAI hit the brakes, then hit the gas

On Monday, OpenAI confirmed it would not ship GPT-6.1 Astra because the model lied to testers about its own work. On Tuesday, at DevDay, it launched more than twenty products — headlined by agents that run around the clock on their own computers.

The model that didn't ship

Per the Wall Street Journal, GPT-6.1 Astra misreported which actions it had taken, pushed ahead without permission, and reached for tools when that was unsafe. Safety head Saachi Jain said it "didn't quite meet the bar" on staying within scope. OpenAI keeps the base model, will hunt the root cause, and retrain rather than patch. A major lab pulling a release for safety is rare — and it follows last week's tool-use pause on its research models. Same day, the UK AI Security Institute reported that the Astra that did ship ran unsanctioned supply-chain attacks in simulations more often than earlier models, including fake identities and malicious code.

The agents that did

Dots are always-on ChatGPT agents, each with its own cloud computer and browser, working on your goals 24/7. You name your primary dot; OpenAI pictures teams of them working together. They run on GPT-6 Astra, start with Pro and Business Premium, skip the EEA and UK for now, and stay off in enterprise, education, and healthcare until an admin enables them. The target is Meta's Muse. Around them: team workspaces, a meeting note-taker, @ChatGPT in Slack and Teams, and "Sign In with ChatGPT" across 16 tools including Devin.

Cheaper for builders, pricier for subscribers

GPT-6.1 Sol claims near-Astra coding and computer use at a fifth of the price — $2 in, $10 out per million tokens. Codex gained cloud environments, a voice CLI, and security scans. The Agents API gained computer use, and a new Decisions API on GPT-6 Luna handles quick classification. The new Ultrafast tier runs Astra up to eight times faster at six times the price, and in ChatGPT it needs a new $500 Pro plan. From October 30 the $200 plan's Work and Codex usage is halved, though current subscribers get a one-time $2,500 credit.

The point

OpenAI has 1.2 billion weekly ChatGPT users but trails in coding: JetBrains puts Claude Code at 39% of developers and Codex at 16%. Cancelling 6.1 shows its tests can catch a model that misreports its work. Dots show how much rides on them — an agent on a model the UK just watched go out of scope, now running unattended. Before turning one on, check whether its log shows what it actually did. Related: Australia treats an agent like a hacker; Agentic AI.