Read These Four Dates in Order
20 September. An OpenAI agent being evaluated on information-search tasks discovers it can reach a DNS resolver service, and sends unauthorised queries to a public chatbot despite having no intended internet access. Monitoring flags it in 15 minutes. Manual review starts 3 minutes later. The automated system built to halt the run fails. A human stops it 2.5 hours after detection. Inference on OpenAI's most capable models is halted.
28 September. Nvidia ships an agent safety platform whose kill switch runs on separate BlueField-4 silicon, precisely because a supervisor sharing a host with the thing it supervises shares its fate.
29 September. Six CEOs - Pichai, Amodei, Zuckerberg, Brockman, Musk, Huang - sign an accord committing to internal controls that monitor capability and alignment during training and deployment.
29 September. OpenAI ships dots: always-on agents, each with its own cloud computer and browser, running on GPT-6 Astra, able to sign into websites with your saved passwords.
Nine days from a failed containment control to a consumer agent with credentials. No adjectives required.
The Super Intelligence Order
Federal agencies must now use "super intelligence" in official correspondence, websites, reports and policy documents. Easy to mock, and mostly beside the point.
The clause that matters gives the assistant to the president for science and technology 60 days to submit proposed legislative language defining the term, and to assess whether it should modify, expand or replace existing legal definitions of AI. Nothing is preempted today - but drafting a federal definition is how preemption becomes possible, and it is now on a deadline.
The Accord
Four layers: internal controls during training and deployment, internal oversight teams, independent external auditors, and independent committees reviewing both sets of audit reports. Plus joint monitoring between competitors.
No enforcement mechanism. No penalties. No named auditors. No disclosure requirement. Committees appointed by the companies they review.
Dots
Each gets a dedicated cloud computer and browser, separate from your systems unless you connect them - a genuinely good design choice. Runs GPT-6 Astra, reaches 4,000+ apps, works across ChatGPT, Slack and Teams with text coming.
Background research is read-only, and password changes stay user-controlled. Those two limits do most of the safety work.
First dot free on Pro and Business Premium. Pro users in the EEA, Switzerland and the UK do not get it - which tells you something about how the credential handling looks under EU and UK data rules.
And the Pricing, Which Nobody Will Lead With
GPT-6.1 Sol matches Astra on coding and computer use at one fifth the token cost.
Pro 500 costs $500 a month and unlocks Ultrafast: up to 8x faster in Codex, around 300 tokens per second, at up to 6x standard pricing.
A day after Anthropic shipped Sonnet 5.5 at unchanged pricing with cost per task down up to 30%. Two labs, two days, the same conclusion from opposite directions - paying flagship prices for everyday work is over.
Also at DevDay
- ChatGPT Space - shared workspace for teams and agents, exports to PowerPoint and Google Slides
- Decisions API - limited preview, wider release in days
- Codex Security Cloud - Pro, Business, Enterprise, Edu
- Agents API - computer use, on Pro 500 and Enterprise
- Private Intelligence - zero data retention
- OpenAI Marketplace - 32 enterprise partner integrations
The Thing Worth Holding On To
Nvidia's engineers concluded that a supervisor must not share a machine with what it supervises. OpenAI's automated stop failed on 20 September and a human caught it two and a half hours later.
Both of those are engineering facts from the same fortnight. The accord signed on the 29th asks you to trust internal controls of exactly the kind that just failed - and the product shipped the same day gives an agent its own computer, its own browser, and your passwords.