Two of the DevDay developer announcements come with restrictions that the keynote did not mention and the documentation does.
The first is Ultrafast, OpenAI’s premium speed tier. The company says it delivers up to eight times faster token generation — 300 tokens per second — in Codex, and up to six times faster in the API. GPT-6 Astra Ultrafast is available now through the API and in ChatGPT Work and Codex on the Pro 500 and Enterprise plans, and OpenAI says GPT-6.1 Sol Ultrafast is coming soon.
Where Ultrafast cannot run
The API guide closes with a sentence worth reading before designing around the tier: “Ultrafast supports US data residency and global processing only. It does not support EU or other non-US regional processing endpoints.”
For a European company that has built on regional processing to satisfy data residency commitments, that makes the fastest tier unavailable without changing where inference happens. It is a practical fork in the road for anyone running latency-sensitive agents under EU data rules, and it is one line in a developer guide rather than a headline.
The guide also sets default token rate limits: 500,000 tokens per minute for API usage tiers one to three, one million for tier four and five million for tier five, and it strongly recommends WebSockets for agentic applications that make many tool calls in quick succession. Note that the guide currently lists preview access for GPT-5.6 Sol, while the DevDay recap refers to GPT-6.1 Sol Ultrafast as coming soon.
The Agents API gets a browser
The second addition is computer use in the Agents API. An agent can navigate websites and operate browser interfaces to test a site, gather information or drive an application through its user interface, with the browser running in an OpenAI-hosted environment. Developers enable it by adding a computer_use tool and switching on a hosted desktop.
The consent model is granular and documented carefully: the browser requires the user’s approval before accessing each new website origin, including public websites, and enabling network access does not approve those requests. The developer’s application has to surface the requested origin and a reason, and collect an approve, deny or cancel decision.
The same page states the limit of that protection in a heading of its own: origin approval does not enforce confirmation before individual actions. Approving a domain approves everything the agent then does on it.
Decisions, and running agents inside AWS
OpenAI also previewed a Decisions API, which narrows a model to a defined set of questions with finite pre-defined answers — classify content, route a request, choose an agent’s next action — with context supplied as text or images. It is in limited preview, with a broader release planned within days. Separately, OpenAI worked with Amazon on Bedrock Managed Agents, which takes the Agents API’s core capabilities and runs OpenAI agents entirely inside AWS.
What to watch
The per-origin approval flow is the right default and it will be the first thing product teams route around, because an agent that stops to ask about every new domain is not an autonomous agent. Watch whether OpenAI ships a blanket-approval option. On Ultrafast, the question is whether EU regional endpoints get the tier at all, or whether speed becomes another thing European deployments trade away for residency.
