OpenAI gives agents a browser and adds an ultrafast tier at DevDay

OpenAI announced a set of new developer tools at its DevDay event on September 29, 2026. The highlights are agents that can use websites in a browser to get work done, a faster response tier called Ultrafast, and the Decisions API, which chooses among predefined answers. A new safety review setup for companies using zero data retention and an agent service running on Amazon's cloud arrived the same day.

The Agents API can now use a browser

The Agents API lets developers add AI agents to their own applications, with OpenAI handling session orchestration, context compaction and recovery. It has been in public beta since September 10. According to OpenAI's changelog, the API now supports computer use: an agent can open websites in an OpenAI-hosted browser and carry out tasks on its own. Developers add the {"type": "computer_use"} tool to a session and choose openai_hosted as the environment type. According to the documentation, the agent asks for approval before it visits each new site, and these approvals, along with any sign-in steps, are handled by the developer's own application. Sign-in details stay outside the content the model sees. According to the DevDay recap, the Agents API also brings Codex's multi-agent, tool search and context compaction capabilities to developers.

Faster responses with Ultrafast

Ultrafast is the fastest service tier in the OpenAI API. According to the company, it generates output up to 8 times faster than Standard mode. For now it works with gpt-6-astra by setting service_tier to "ultrafast" in the request, and support for GPT-6.1 Sol is described as coming soon. OpenAI recommends a persistent WebSocket connection for agent applications that make many tool calls, so the speed gain is not lost to network overhead. The tier runs only with global processing and US data residency; EU and other regional data residency options are not supported. Ultrafast costs more than the standard tier, and current prices are on OpenAI's website.

The Decisions API for tasks that need a choice

The Decisions API focuses OpenAI's Luna model on questions with predefined answers. Sorting a request into the right category, routing it to the right place or picking an agent's next step are typical examples. Text or images can be given as context. The API is in limited preview for now, and OpenAI says wider access will follow in the coming days.

Private safety review with zero data retention

ZDR with Private Safety Processing is designed for customers approved for zero data retention (ZDR). Prompts and responses selected for automated safety review are not kept by OpenAI; they are encrypted and written to the customer's own cloud storage in AWS S3, Azure Blob or Google Cloud Storage. Each record is encrypted twice: an inner layer with OpenAI's HPKE encryption and an outer layer added through the customer's own key management (EKM). Records are decrypted only in a hardware-attested review environment with human access disabled, and only limited safety signals leave that environment. Records expire after 30 days. If the customer revokes its key authorization, the records can no longer be decrypted; according to the documentation, this does not delete them or undo completed reviews. According to the DevDay recap, a preview of a new approach called Private Inference is coming this fall.

OpenAI agents on Amazon's cloud

AWS opened Amazon Bedrock Managed Agents in preview the same day. Built jointly by AWS and OpenAI, the service runs on a version of the OpenAI Agents API adapted for AWS and is available for now in the US East (N. Virginia), US West (Oregon) and US East (Ohio) regions. Durable sessions let an agent pick up where it left off, and agents can connect to MCP servers. Each agent runs with its own IAM role, human approval can be required before consequential actions, and supported activity is recorded in CloudTrail.

The Agents API and computer use are still in beta, while the Decisions API and Bedrock Managed Agents are in preview. According to the documentation, site access approval does not guarantee a separate confirmation before irreversible actions such as purchases or deletions, so applications need their own approval step for those. Browser screenshots can contain sensitive information and should be shown only to authorized people. The lack of EU data residency for Ultrafast is an important limit for companies working with data in Europe. To see how these tools hold up in real work, it is worth following evaluations from independent developers and researchers.

Sources: OpenAI: DevDay 2026 recap, OpenAI API changelog, OpenAI: Agents API computer use, OpenAI: Ultrafast mode, OpenAI: ZDR with Private Safety Processing, AWS: Amazon Bedrock Managed Agents preview.