Decision API from the original description and OpenAI Codex Cloud: decision automation and long-running tasks

The original description presents Decision API as an interface for automatically choosing from a predefined set of options. The provided OpenAI materials do not confirm the launch of a product by that name. Therefore, the scenario below is treated as a description from the original brief, not as a confirmed OpenAI feature. OpenAI’s help documentation describes Codex Cloud as running development tasks in a cloud environment, where work can continue even when the computer is turned off.

Decision API: how limited choice differs from a chat prompt

According to the original description, a developer specifies a set of permitted answers, such as “approve” or “reject,” and the system selects one. The provided OpenAI materials do not confirm that an API by this name exists or works this way. Choosing from a predefined set of options is described here only as a hypothetical scenario.

This scenario could suit repetitive operations if a business rule can be expressed through a limited set of outcomes. For example, an application could request an additional review when there is not enough input data to make a choice. This is an example of a possible implementation, not a confirmed product capability: the accompanying materials do not provide information about its interface or technical requirements.

How to specify permitted options

In the hypothetical scenario, a fixed list of answers constrains the result but does not, by itself, make it possible to assess whether the result is correct. Suppose the system must choose “approve” or “reject” even though the input data is incomplete or contradictory. The result will still be one of those two decisions, but that does not prove it is correct.

When designing such a system, a developer should determine what data the model receives, what each option means, and what to do when there is uncertainty. If there is no separate outcome such as “refer to a person,” the application can be configured to detect insufficient data and stop automatic execution. This is a design recommendation, not a requirement or feature confirmed by the OpenAI materials.

• Specify the permitted outcomes and describe their meanings unambiguously.
• Test the system with examples involving incomplete and contradictory data.
• Separate choosing an answer from carrying out an action, such as sending a payment or changing a record.
• Keep manual review available wherever an error could have significant consequences.

Codex Cloud: tasks continue when your computer is off

Codex Cloud runs development tasks on OpenAI-managed computers and uses prepared cloud environments with project repositories, tools, and dependencies. Each task gets its own isolated workspace. OpenAI documentation states that a cloud task can continue while the user’s computer is turned off.

Before launching a task, you need to create and publish a cloud environment through the web interface or desktop app. You can then start or continue a task on a computer, in a browser, or on a smartphone. To launch a task from a mobile device, first configure and publish the environment on a computer or in the web version.

Running a task in the cloud does not mean remotely controlling a task on your own computer. For Codex to work on a user’s device, remote access depends on whether the device is turned on and whether its settings allow the connection. In Codex Cloud, the task runs on cloud infrastructure.

Collaboration between people and AI

Codex lets you work on development tasks across different devices. The provided materials do not substantiate the broader claim that tasks can be handed off between team members and agents, so no such capabilities are asserted here. Specific features depend on the client, plan, and workspace settings. In managed workspaces, the administrator separately controls access to Codex Cloud. Permission to run cloud tasks does not automatically mean that local work in Codex is allowed.

When handing off a task, provide context: the repository, the goal of the change, constraints, and verification criteria. If the result includes code changes, the team needs to review them and run the checks used by the project. An agent completing a task does not replace a review.

Generation speed and the Pro plan

The original description mentions HGPT Work mode, with speeds of up to 300 tokens per second, and a Pro subscription for 500 dollars per month. The provided OpenAI materials do not substantiate this name, speed figure, or access terms. They refer to a different product: GPT-6 Astra in Ultrafast mode, available on the Pro 500 plan and in certain Enterprise and Edu workspaces.

According to OpenAI’s help documentation, Ultrafast uses the included Work and Codex allowance, and credits may be used after it runs out. Access depends on the plan and workspace settings. Simply purchasing credits on Pro 100 or Pro 200 does not unlock Ultrafast. The speed of up to 300 tokens per second and the claim that it is “about eight times faster” are not substantiated by the available materials, so they cannot be treated as established specifications.

Cost and availability of Codex Cloud

Codex is available on ChatGPT plans, but separate terms apply to Codex Cloud. OpenAI’s help documentation lists eligible Plus, Pro, Business, Enterprise, Healthcare, and Education accounts; access also depends on workspace settings and feature availability. Free and Go plans do not include Codex Cloud. Usage limits depend on the plan, model, and task itself.

The Codex Cloud description does not list a separate charge for a standard cloud virtual machine. Model usage counts toward regular Codex limits and, where applicable, credits or billing. A long-running task may use more resources than a short prompt. Check the exact terms in your account and workspace settings.

Risks of automating decisions with fixed options

In the described scenario, limited choice sets the response format, but it does not guarantee that the model or input data is free of errors. If every case is reduced to two options, an exception may receive an inappropriate label. For financial, employment, and other consequential decisions, it is worth having clear review rules, logging, and the option of human reconsideration.

Caution is also needed when Codex works with a repository and connected tools. The agent’s access is determined by the user’s permissions and workspace settings. Actions that require confirmation should go through the designated review process. Do not give the agent more privileges than it needs for a specific task.

What is confirmed about OpenAI’s new tools

• Codex Cloud runs development tasks in cloud environments, and they can continue while the user’s computer is turned off.
• Codex cloud tasks are accessible from a computer, the web version, and a smartphone, subject to access requirements and prior environment setup.
• Codex Cloud usage is affected by plans, limits, administrator permissions, and credits.
• The provided OpenAI documents do not substantiate Decision API, HGPT Work mode, or the figure of 300 tokens per second.

Frequently asked questions about Decision API and Codex Cloud

Can Decision API be considered a launched OpenAI product?

The provided materials do not support that claim: they contain no confirmation that an API by this name has launched. The description of choosing from a fixed set of options should be treated as a claim in the original brief, not as a feature confirmed by the documentation.

Does your computer need to stay on for Codex Cloud?

No. Codex Cloud runs tasks on OpenAI cloud computers, so a task can continue when the user’s computer is turned off. This differs from remotely accessing a task that is running on your own device.

Is a speed of 300 tokens per second confirmed?

No, the accompanying materials do not substantiate this figure. They describe Ultrafast mode for GPT-6 Astra, but do not provide a speed figure.