On September 29, 2026, at its DevDay developer conference in San Francisco, OpenAI announced that the Agents API has officially entered public beta, and at the same event introduced Computer Use capability integration, the Decisions API, an Ultrafast speed tier, and AWS Bedrock managed integration. This is OpenAI's most intensive release to date at the agent infrastructure layer, and a clear signal that AI agents are shifting from assisting with text generation to autonomously executing tasks.
The Agents API is now open to all developers, with no waitlist application required. Under the hood it runs on the same open-source harness that powers Codex and ChatGPT Work, hosted and maintained by OpenAI. With the addition of Computer Use, agents can click, fill in forms, and operate any software interface through OpenAI-hosted browsers; billing is calculated by tokens, tool calls, and container usage, and Computer Use itself carries no additional charge. The Ultrafast tier raises API token generation speed by up to 6x, and Codex by up to 8x to reach 300 tokens per second, but at the cost of 6x standard-tier pricing: for GPT-6 Astra, output tokens are priced at $300 per million.
From Four Concepts to a Single API Call
The technical design of the Agents API is organized around four core concepts: Agent (a configuration set of models, instructions, tools, and MCP servers), Environment (a sandbox in which the agent accesses files, loads skills, and executes commands), Session (a persistently running agent instance that can receive tasks and continuously respond to input), and Events and Items (inputs sent to the agent and outputs the agent produces). Developers can use a single API call to create an incident investigation agent that dispatches three sub-agents in parallel, calls external MCP servers, and writes results to a specified working directory — the entire configuration is completed in under 30 lines of JavaScript code.
The Agents API supports three sandbox modes: OpenAI-hosted sandbox (the same isolated infrastructure used by Codex and ChatGPT), self-hosted mode (running codex exec-server in your own environment and registering outward over WebSocket, with all connections outbound), and partner sandboxes (nine in total, including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel). Enterprises can choose a mode in which data never leaves their own infrastructure, but Agents API data is currently processed only within the United States and does not support zero data retention agreements.
This limitation matters significantly on the compliance dimension. For enterprises bound by EU GDPR, or financial and healthcare institutions with strict data localization requirements, the data-stays-in-the-US clause is directly a procurement barrier, and the value of the AWS Bedrock integration is constrained accordingly.
Decisions API: The Underestimated Speed Difference
The Decisions API released the same day targets classification, routing, and single-step decision scenarios. A decision completed on GPT-6 Luna has a latency of about 150 milliseconds, whereas calling GPT-6 Luna through the standard API takes about 1.6 seconds. The Decisions API is currently in limited preview and not yet generally available.
For agent orchestration scenarios, 150 milliseconds matters: in agent workflows involving multi-step scheduling, if every routing decision requires waiting more than a second, the latency of the entire task chain compounds quickly in complex scenarios. The Decisions API compresses such decisions to the 100-millisecond level, theoretically allowing agents to complete more loop iterations within the same time window. For enterprise scenarios requiring high-frequency decisions — real-time customer service, dynamic form filling, multi-system workflow orchestration — this number has more direct engineering value than Computer Use.
Ultrafast: What Is the Speed Premium Worth?
Ultrafast is the most controversial part of this release on pricing. At $300 per million output tokens, it is 30 times the output price of GPT-6.1 Sol. OpenAI positions it as a premium tier for when speed matters most, available on the Enterprise plan and the newly launched Pro 500 subscription.
A 6x or 8x increase in token generation speed does not equal a 6x or 8x reduction in task completion time. During execution, agents spend a great deal of time on tool calls, code execution, file reads and writes, and network requests — none of which are affected by token generation speed. The real-world speedup depends heavily on task structure. If the bottleneck is tool waiting rather than model inference, Ultrafast's premium is hard to realize. For developers, measuring the share of actual task time spent on token generation with the Trace tool before paying for Ultrafast is the rational path to a decision.
GPT-6.1 Sol offers a clearer value proposition: input priced at $2 per million tokens, output at $10 per million, and cached input down to $0.10 per million — performance close to GPT-6 Astra at less than a quarter of the price. For most enterprise agent workloads, Sol's cost structure is more sustainable than the Astra Ultrafast combination.
AWS Integration: Enterprise Lock-In or Ecosystem Opening?
Native integration of AWS Bedrock Managed Agents means enterprises already operating within the AWS ecosystem can orchestrate OpenAI agents directly through Bedrock, without setting up a separate presence on the OpenAI platform. Agent compute can run in the OpenAI-hosted sandbox, on the enterprise's own infrastructure, or in one of nine partner sandboxes, and AWS resources can be accessed directly by agents.
A Marketplace with 32 launch partners was also introduced, including Figma, Adobe, Salesforce, and ServiceNow. With Computer Use and this partner list, agents do not need to wait for Figma or Salesforce to open dedicated APIs — they can operate the interface directly to complete tasks. This is especially important for enterprises with legacy internal systems that lack modern API interfaces.
Comparison with Existing Capabilities
Computer Use is not a direction OpenAI invented. Anthropic integrated Computer Use capability into Claude 3.5 Sonnet at the end of 2024 and opened the API to developers. OpenAI is now folding similar capability into the unified framework of the Agents API, and Computer Use itself carries no extra charge. Because the two have different billing structures for containers and tool calls, a direct comparison of total cost requires a specific workload to reach a conclusion.
Computer Use went live in Codex and ChatGPT on macOS as early as April this year, and expanded to Windows 11 in May. What DevDay released is the API interface that exposes this capability to third-party developers, not the first appearance of the capability itself.
Strategic Assessment: Where to Look for the Next Signal
The following is analytical judgment, not established fact.
The combination of releases at this DevDay reveals a clear bet by OpenAI in the enterprise market: make agent infrastructure a platform, not just a model. The four-concept architecture of the Agents API, nine sandbox partners, 32 Marketplace members, native AWS Bedrock integration — put together, this is closer to the logic of building the AWS Lambda ecosystem than to a simple API product launch.
Whether this bet pays off has three most critical verification signals: First, whether the data-stays-in-the-US restriction and the lack of zero data retention support will be lifted in the coming quarters, which directly determines whether large enterprises in Europe and Asia can enter. Second, the pricing of the Decisions API after the limited preview period ends — if reasonably priced, its impact on agent orchestration costs could be more substantial than Ultrafast's. Third, the stability and operational cost of self-hosted sandboxes: when enterprises choose to run codex exec-server on their own infrastructure, the depth of OpenAI's support for that path will determine whether it can truly replace enterprises' in-house agent scheduling layers.
Ultrafast's steep premium looks more like a niche product for a small number of extremely latency-sensitive scenarios than a mainstream choice for enterprise-scale deployment. What really shapes the enterprise agent cost curve is the combination of the Sol pricing system and the Agents API hosted sandbox. Most enterprises' agent workloads will ultimately run on Sol rather than Astra Ultrafast, and will choose the OpenAI-hosted sandbox to avoid self-managed operations costs. Whether this combination can achieve sufficient economies of scale within six months is the core indicator for judging whether this DevDay truly constitutes an enterprise AI inflection point.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接