OpenAI is rolling out Astra to ChatGPT Work, Codex, and API, emphasizing enterprise integration, stronger security, and workflow automation for
OpenAI announced that its latest model, Astra, is being rolled out to ChatGPT Work, Codex, and the API, enabling enterprises and developers to use the model directly in workflows for tasks such as coding, business documents, spreadsheets, presentations, and email, even when some applications do not themselves have an API.
Astra also becomes the foundation model for ChatGPT Work and Codex, and supports enterprise application integration.
This update also includes new enterprise plugins for the ChatGPT Desktop app, covering Oracle Analytics, Power BI, Navan, and Avalara, with the goal of bringing enterprise data and analytics workflows into ChatGPT's operational scope.
OpenAI also emphasized that Astra is the first model to reach the Preparedness Framework's Critical level for cybersecurity capability, and is also the model most able to follow user intent.
In terms of pricing, Astra's input and output fees start at $10/$50 per million tokens.
On the other hand, because demand surged after Astra's launch, OpenAI temporarily paused new subscriptions and upgrades for its monthly $200 Pro plan to ensure that existing users could continue using Astra.
Astra's product positioning is clear: it is not simply about improving natural language generation, but about evolving into a high-end execution model for computer use, web browsing, professional work, software development, cybersecurity, and scientific tasks.
This means the model's value lies not only in generating text, but in whether it can complete more complex, context-dependent work with fewer tokens and fewer retries.
From an enterprise productivity perspective, Astra is designed to participate directly in existing toolchains, including ChatGPT Work and Codex, allowing users to complete development or business tasks through the applications they already use.
This kind of integration lowers workflow switching costs and makes the model closer to real enterprise operations rather than just a general conversational assistant.
OpenAI's published evaluation results indicate that Astra outperforms competitors in nearly every professional work and coding test, including Terminal-Bench 4.0 and the Artificial Analysis Intelligence Index.
Technically, this suggests its advantages may come from stronger task-following ability, more stable tool-calling performance, and first-draft output quality that is closer to team standards.
Security is one of the main highlights of this update.
OpenAI said that in internal safety tests for leaking confidential information, improperly exposing dashboard data, or deleting data, Astra produced unintended outcomes 89% less often than the previous-generation model, and 74.7% less often than another model.
This indicates improved behavioral stability in sensitive tasks, while also showing that enterprise deployment should still treat the model as a controlled tool rather than a fully trusted autonomous operator.
In addition, ChatGPT Work and Codex now include confirmation rules and other safety checks that can require ChatGPT to obtain permission before taking important actions, and can automatically check for unsafe or unauthorized tool calls.
Enterprise administrators can also control which websites, desktop applications, upload and download actions, and browsing history Astra can access.
These controls show that Astra's deployment model has moved beyond a simple model service and into the realm of enterprise governance and permission boundaries.
In practical cases, OpenAI has already used Astra for marketing video production and engineering debugging.
The marketing team used Astra and Codex to turn three hours of multi-camera footage into a promotional video, which received 550,000 views within four days; the engineering team used it to identify a memory allocation bottleneck that slowed Codex sessions in a test environment, and after switching allocators, latency dropped by 25x, though peak memory usage increased by about 30%.
These cases suggest Astra is well suited to tasks that require step-by-step reasoning, tool collaboration, and trade-off decisions.
For enterprises, Astra's impact is first reflected in faster knowledge-work workflows.
Presentations, documents, spreadsheets, email, and software development can all be brought into a single model-driven process, creating clear efficiency gains for large volumes of repetitive professional work.
If an enterprise lacks a complete API or system integration capability, Astra can lower the adoption barrier by allowing AI to intervene directly at the desktop-operation layer.
For security teams, Astra represents a higher-level dual effect.
On the one hand, it can be used for cybersecurity and scientific tasks, and its model safety is more stable than previous generations.
On the other hand, when the model is authorized to operate websites, downloads, uploads, desktop applications, or enterprise data, any mistake, privilege escalation, or misjudgment could expand into risks of data leakage, unauthorized actions, or operational disruption.
For IT governance, this type of model is no longer just a SaaS feature, but a core system that must be brought under permissions, auditing, data classification, and change management.
If enterprise administrators adopt a phased approach that starts with partial access and expands gradually, that means real-world implementation must first verify whether minimum necessary permissions, tool allowlists, and confirmation rules are sufficient to cover high-risk workflows.
For cost and resource management, although Astra claims to complete tasks with fewer tokens and fewer repeats, its pricing model and high usage demand also mean enterprises must reassess inference costs.
Especially when the model is broadly embedded in daily workflows, token consumption, approval counts, and plugin data-read frequency will directly affect total cost of ownership.
When deploying Astra, enterprises should treat it as a high-privilege AI work agent rather than a simple chat tool.
The top priority is least privilege: avoid opening too many websites, desktop applications, and data sources at once, and start with a limited scope before gradually expanding access.
For tasks that involve sensitive data, financial information, customer data, or core system settings, confirmation rules and human review should be enabled.
Any action involving deletion, downloading, batch changes, or external transfer should require explicit confirmation to prevent the model from causing irreversible effects in the toolchain.
Enterprises should also establish mechanisms to detect unsafe or unauthorized tool calls and retain AI operation logs for audit purposes.
If plugins integrate with systems such as Oracle Analytics, Power BI, Navan, or Avalara, the data access scope, output masking, and permission inheritance relationships should be further reviewed to prevent cross-system data spread.
At the management level, it is recommended to incorporate Astra usage into formal security policies and change processes, including model usage scope, allowed tasks, exception handling, log retention, and incident response procedures.
If the model is used for software development, its output should go through code review, test validation, and pre-deployment review to avoid sending high-quality but unverified first-draft output directly into production.