OpenAI introduces GPT-6 Astra for work

OpenAI has introduced GPT-6 Astra, a model designed to perform professional tasks directly inside the tools companies already use. It is available in ChatGPT Work, Codex, and the API, although enterprise access is disabled by default.

The main difference is that Astra does more than answer questions or generate text. It can write code, browse websites, and operate desktop applications, even when those applications do not offer an API, meaning a technical connection designed for other programs to control them.

What Astra can do

According to OpenAI, the model is designed for work that requires more than a single answer: reviewing information, making decisions according to instructions, and completing tasks across multiple tools.

Its announced uses include:

• Analyzing financial statements and detecting discrepancies.
• Optimizing graphics card usage for AI workloads.
• Creating presentations that follow a company’s voice, templates, and visual standards.
• Programming and working inside common applications, even when they do not have an API.
• Performing tasks related to science, cybersecurity, and web browsing.

This removes one of the usual barriers to using AI in a company. Instead of preparing all the data, redesigning processes, and building custom integrations, a team could give it controlled access to the same applications it already uses.

OpenAI says it tested Astra internally before launch. In one case, its engineering team identified a bottleneck in Codex session memory. After changing the memory allocation system, response latency fell 25 times, although maximum memory usage increased by approximately 30%.

More work for every dollar

The company says Astra completes tasks with fewer tokens, which are the units of text a model uses to process information, and needs fewer repeated attempts. In practice, that can mean fewer corrections and a lower cost per task.

The announced price starts at 10 dollars per million input tokens and 50 dollars per million output tokens. The final cost will depend on how much content the model processes and how many actions it needs to perform.

Controls to prevent unauthorized actions

Giving an AI access to enterprise systems also increases the risk that it could share information, modify data, or take the wrong action. To limit that risk, Astra requires confirmations before important operations and automatically reviews potentially dangerous or unauthorized calls.

Administrators can restrict:

• Allowed websites and desktop programs.
• File uploads and downloads.
• Browsing history.
• The scope of access for each team or user.

In an internal OpenAI evaluation focused on safe computer use, Astra was 89% less likely to produce undesired outcomes than GPT-5.6 Sol and 74.7% less likely than Claude Fable 5.1. The test simulated situations such as exposing confidential information, sharing a dashboard with too many people, or deleting data. These are results from an internal benchmark, not a guarantee that the model will never make a mistake.

OpenAI also says Astra is its first model to reach the critical capability threshold in the cybersecurity section of its preparedness framework. That is why it says it has strengthened its defenses against malicious use, attempts to bypass protections, and unauthorized actions.

ChatGPT Desktop also adds enterprise plugins from Oracle Analytics, Power BI, Navan, and Avalara. Some organizations that meet the requirements will also be able to use zero data retention on certain API endpoints.

For you, the important change is not just that Astra writes better. It can act inside work tools, which makes permissions and confirmations just as important as the model’s capabilities. What needs watching now is how these controls perform in real-world situations and how many companies decide to expand access beyond limited configurations.