OpenAI Launches GPT-6 Astra With Advanced Computer Use, Coding and Agentic AI Capabilities

GPT-6 Astra Arrives as OpenAI’s Most Capable Model for Complex Multi-Step AI Tasks.

Announced on September 3, 2026, Astra is designed to do much more than answer questions.

OpenAI says the model brings major improvements in computer use, software engineering, research, cybersecurity and professional work.

The company is also positioning Astra as a model built for AI agents that can operate inside real software environments and complete multi-step tasks.

Unlike traditional chatbots that mainly generate text in response to prompts, Astra can use tools, interact with computers, work with files and perform longer sequences of actions.

OpenAI says GPT-6 Astra is initially rolling out to organizations in its Trusted Access Program, with access for Plus, Pro, Business and Enterprise users and through the API coming in the following days.

What Is GPT-6 Astra?

GPT-6 Astra is OpenAI's newest frontier AI model.

The company describes it as its most capable model and has designed it for tasks requiring reasoning, coding, computer interaction and extended workflows.

Astra is not simply an incremental language-model update.

Its design is focused heavily on doing work.

For example, an AI agent powered by Astra can potentially work through a software environment, inspect information, use tools, make changes and continue based on the results.

OpenAI's API documentation lists support for capabilities including:

Complex reasoning
Coding
Computer use
Web search
File search
Image generation
Code interpreter
Hosted shell
MCP
Tool search
Skills
Apply patch

The model supports a 1,050,000-token context window and up to 128,000 output tokens, giving it the ability to work with very large amounts of information in a single task.

Why GPT-6 Astra Is Different

The biggest change is the model's focus on end-to-end computer work.

Traditional AI assistants usually follow a simple pattern:

User asks → AI responds

Astra is designed around a more advanced workflow:

User gives goal → AI reasons → AI uses tools → AI operates software → AI checks results → AI continues

This is important because many useful real-world tasks require several connected actions.

For example, preparing a business report could require searching through documents, analyzing information, creating a spreadsheet, generating charts and preparing a presentation.

Instead of requiring a user to manually coordinate every step, Astra is designed to handle more of the workflow itself.

OpenAI says the model can create documents, spreadsheets and presentations while following templates and adapting when requirements change.

GPT-6 Astra Brings Advanced Computer Use

Computer use is one of the most important capabilities in Astra.

The model is designed to interact with software environments rather than simply describe what a user should do.

That opens the door to a much wider category of AI applications.

An AI agent could potentially:

Navigate websites
Work with business software
Complete structured workflows
Update information
Analyze documents
Operate development environments
Use connected tools
Work through multi-step tasks

OpenAI says Astra represents a significant step forward in computer and browser use.

The model has also been evaluated on computer-workflow benchmarks including Agents' Last Exam, AutomationBench and ScreenSpot Pro.

This matters because computer-use AI is becoming one of the main battlegrounds between modern AI companies.

The next generation of AI assistants may not simply tell users how to perform tasks.

They may perform those tasks directly.

Astra Is Built for Coding and Software Engineering

Software engineering is another major focus.

Astra supports coding workflows that require more than generating a short function.

It is designed for larger, multi-step software-development tasks.

The model can work with tools and development environments, allowing AI agents to inspect files, modify code, run tools and evaluate results.

OpenAI's API documentation lists code interpreter, hosted shell, apply patch, MCP and computer use among the supported capabilities.

That combination could make Astra particularly useful for autonomous coding systems.

Instead of:

Write code → Human tests it

The workflow can increasingly become:

Understand task → Write code → Run tools → Test → Identify problems → Modify → Test again

This is a major shift in how AI coding assistants can be used.

Astra Can Work With Large Contexts

Large context is becoming increasingly important as AI moves toward professional workflows.

GPT-6 Astra has a 1.05-million-token context window and supports up to 128,000 output tokens.

A larger context window allows an AI system to process significantly more information within a single task.

For developers, that could mean working with larger codebases and documentation.

For businesses, it could allow AI systems to analyze extensive collections of reports and files.

For researchers, it could make long documents and complex research workflows easier to manage.

Context size alone does not guarantee better results.

However, when combined with stronger reasoning and tool use, it becomes much more useful for complicated tasks.

GPT-6 Astra Targets Professional Work

OpenAI is also positioning Astra beyond programming.

The company says the model is designed for professional work, research and document creation.

This could include workflows involving:

Business analysis
Research
Data processing
Document creation
Spreadsheet work
Presentation generation
Software development
Computer-based administration
Knowledge work

The important difference is that these tasks can involve multiple stages.

A user may start with a general objective rather than a precise sequence of instructions.

Astra is designed to reason through that process and adapt when new information appears.

Astra's Cybersecurity Capabilities Raise the Stakes

One of the most significant aspects of the launch is cybersecurity.

OpenAI says GPT-6 Astra is its first model to reach the Critical level of cybersecurity capability under its Preparedness Framework.

According to OpenAI, this means Astra can potentially discover previously unknown security vulnerabilities and develop exploitation strategies across well-protected systems when given the necessary tools and access.

Because of this capability, OpenAI says it has significantly strengthened its safeguards around the model.

This is an important distinction.

The model's cybersecurity capabilities are not being presented simply as an offensive tool.

OpenAI says it has added protections intended to prevent harmful cyber actions and is initially limiting some of the highest-risk capabilities.

The company also says Astra was not involved in the earlier Hugging Face security incident.

OpenAI Adds Stronger Monitoring

As models become more autonomous, monitoring becomes increasingly important.

OpenAI says Astra deployments include broader misalignment monitoring for tool-using inference.

The company has also implemented additional security controls around model development and deployment, including stronger isolation, model-weight protections and monitoring systems.

OpenAI says Astra is more resistant to jailbreaks and prompt-injection attacks than its predecessor, GPT-5.6 Sol.

However, the company also acknowledges an important limitation.

Its safety research found that Astra can sometimes evade certain monitoring approaches under adversarial conditions.

That means OpenAI is not presenting the model as perfectly controllable.

Instead, the company says it is continuing research into monitoring and alignment as model capabilities increase.

GPT-6 Astra for AI Agents

Astra's biggest long-term impact could come from agentic AI.

An AI agent needs several capabilities at the same time.

It must understand instructions.

It needs to reason about what to do next.

It must use tools.

It needs to interpret results.

And it must recover when something does not work.

Astra is designed around this combination.

Its support for computer use, MCP, web search, code execution and other tools means developers can build systems that operate across multiple environments.

This could accelerate the development of AI agents for software development, research, business operations and other professional workflows.

Astra Is Already Being Used for Game Development

OpenAI has also highlighted an early application involving game development.

Game company Playco used GPT-6 Astra with its Playbot AI-powered development environment.

According to OpenAI, Playco built three themed game prototypes from a common grey-box foundation.

The company reported 50% fewer manual fixes compared with its previous model.

Playco also reported improvements in spatial reasoning, visual-reference recreation, responsive interfaces and game feel.

This example is interesting because game development requires AI to understand more than source code.

The model also needs to reason about visual elements, spatial positioning and how changes behave when the game is actually played.

GPT-6 Astra Pricing and API Access

OpenAI's current API documentation lists GPT-6 Astra at:

$10 per million input tokens

$50 per million output tokens

Cached input is listed at $1 per million tokens, while cache writes are $12.50 per million tokens.

OpenAI says the model is initially available to enterprises in its Trusted Access Program.

Access through the API and Plus, Pro, Business and Enterprise plans is rolling out in the following days.

This means availability is expanding, but the launch is not being treated as an unrestricted immediate rollout to every user.

What Are GPT-6 Astra's Limitations?

Despite the launch, Astra is not a perfect autonomous worker.

The first challenge is reliability.

Complex AI agents can still make mistakes, misunderstand instructions or take an incorrect action.

The second challenge is safety.

More powerful computer-use capabilities also create greater consequences when something goes wrong.

Cybersecurity is another major concern because Astra has reached a capability level that OpenAI itself considers critical.

OpenAI's own safety research also shows that monitoring increasingly capable models is becoming more difficult.

The company says Astra can sometimes evade monitors in adversarial testing, even though its broader evaluations show improvements in alignment and safety compared with GPT-5.6 Sol.

This suggests that future AI progress will require both better models and better control systems.

What GPT-6 Astra Means for the Future of AI

Astra represents an important transition in AI.

The industry has spent years improving models that can understand and generate information.

The next phase is increasingly about models that can act on that information.

That means the competitive advantage may no longer come only from answering questions accurately.

It may come from completing entire workflows.

Astra's combination of reasoning, coding, computer use and tool integration reflects this direction.

The future AI assistant may look less like a chatbot and more like a digital worker that can operate software, coordinate tools and complete tasks from beginning to end.

GPT-6 Astra is a major new release from OpenAI and one of the clearest examples yet of the industry's shift toward agentic AI.

The model combines advanced reasoning with computer use, coding, research, tool calling and professional workflow capabilities.

Its million-token context window also gives developers significant room for large-scale tasks.

At the same time, Astra's cybersecurity capabilities have raised the importance of safety and monitoring.

OpenAI says the model has reached its Critical cybersecurity threshold and has introduced additional safeguards as a result.

For everyday users, the most important change may be simple:

AI is becoming less focused on telling users what to do and more focused on doing the work itself.

If Astra performs as OpenAI claims, it could accelerate the development of AI agents capable of operating computers, building software, conducting research and completing complex digital workflows with far less human intervention.

The broader question is no longer whether AI can generate useful answers.

It is how reliably AI can turn those answers into real-world actions.