OpenAI introduced GPT-6 Astra on September 3, 2026, with a promise that goes beyond generating better answers. The company wants its new model to perform substantial professional work: operating software, writing and testing code, researching complex topics, and producing documents that follow an organization’s requirements.
The difference is potentially significant. Instead of asking an AI assistant to explain how to complete a task, users can increasingly delegate parts of the task itself. Through ChatGPT Work and Codex, Astra can combine reasoning with tools and computer access to carry out activities that previously required extensive manual intervention.
But greater autonomy also creates new questions. How reliably can the model execute lengthy assignments? How much supervision does it need? And do its improved capabilities justify the additional cost?
What Is GPT-6 Astra?
GPT-6 Astra is OpenAI’s flagship model in the GPT-6 family. According to the company’s official announcement, its development combines advances in pre-training, reinforcement learning and alignment.
Its principal areas of improvement include software engineering, computer use, web research, scientific reasoning and professional tasks requiring several consecutive actions.
Astra is designed to work through complicated assignments while retaining the original objective. For example, a user might ask it to research an industry, analyze several documents and turn the findings into a presentation using an existing company template.
The important development is not that AI can generate a presentation. Earlier models could already produce office documents. Astra is intended to handle more of the surrounding workflow, from collecting information and interacting with applications to incorporating additional instructions without losing track of the original assignment.
From Chatbot to Computer-Using Agent
One of Astra’s most consequential capabilities is computer use.
Traditional language models primarily interact with the world through text and specially designed tools. Computer-using agents can go further by navigating graphical interfaces, entering information, opening documents and interacting with applications.
OpenAI says Astra can perform activities such as updating customer records, filling in forms, organizing calendars, researching information and testing websites. It can also work with applications that do not provide a dedicated API.
Consider a hypothetical business workflow. A project manager needs to prepare a monthly report using information from a CRM, several spreadsheets and internal documents.
An AI assistant might previously have summarized the information after someone collected and uploaded it. With suitable access, an agent powered by Astra could potentially retrieve the relevant material, compare figures, identify missing information and prepare a report using an approved template.
That does not mean the report should be accepted without review. The agent’s permissions, source selection, calculations and final output still need appropriate checks.
The distinction between a model’s intelligence and its access to external systems is important. Our guide to the Model Context Protocol explains how AI applications connect to information and tools, and why those connections require carefully defined permissions.
What Do Astra’s Benchmarks Actually Show?
OpenAI has published extensive evaluations comparing Astra with GPT-5.6 Sol and several competing models.
Three results illustrate the changes in the areas most relevant to professional use.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| OSWorld 2.0 — Computer use | 72.6% | 65.7% |
| Terminal-Bench 4.0 — Coding and terminal tasks | 57.9% | 37.3% |
| AutomationBench — Professional automation | 41.4% | 18.1% |
Source: OpenAI’s published GPT-6 Astra evaluations. Results depend on the benchmark version, execution environment and testing settings.
The particularly large improvement on Terminal-Bench suggests that Astra can handle more demanding technical workflows than its predecessor under the reported conditions.
On OSWorld, OpenAI also reports that Astra completed evaluated tasks in substantially less simulated time. This matters because an autonomous agent’s usefulness depends not only on whether it eventually reaches the correct answer, but also on the time and resources required.
However, the figures should not be interpreted as universal success rates. A benchmark result does not establish that Astra will complete the same proportion of tasks in a particular company’s software environment.
Users also need to consider the surrounding agent system, available tools, reasoning settings and verification requirements. Independent testing on representative workloads remains essential.
For more on interpreting these figures, see our article Top of the AI Benchmarks: Is It the Best Model for You?.
Codex: Better Support for Long Programming Tasks
Software development is another major focus of Astra.
Beyond generating code, advanced coding agents need to understand existing projects, inspect dependencies, run tests, investigate failures and avoid introducing unrelated changes.
One particularly interesting improvement concerns long development sessions.
When a conversation or coding task exceeds a model’s context window, previous systems often compress earlier information into summaries. This process can lose important details, including failed debugging attempts or architectural decisions.
OpenAI has introduced an experimental context-management approach for Astra in Codex. The system can preserve notes across context windows and search earlier context when it needs to retrieve requirements, previous test results or other relevant information.
This could be valuable for demanding tasks such as migrating an application to a newer programming language, refactoring a large codebase or investigating a complicated software defect.
However, better context management does not eliminate the need for version control, automated tests and human review. A development agent should still demonstrate that its changes work and respect the project’s constraints.
Professional Documents: The Quality of the Finished Result Matters
OpenAI also emphasizes Astra’s ability to produce documents, spreadsheets and presentations that follow existing templates.
The challenge here is not simply generating attractive content. A professional deliverable must preserve the relevant facts, respect formatting requirements and remain usable when transferred into the organization’s normal workflow.
Astra is designed to maintain these requirements as the task evolves. Users can introduce new constraints, ask questions about intermediate results or change direction without necessarily restarting the assignment.
This positions ChatGPT Work more directly within everyday business activities, from financial analysis and project reporting to research and document preparation.
Competition in this area is developing quickly. Anthropic is also expanding its document-oriented workspace, as discussed in our analysis of Claude Docs and Slides.
For organizations evaluating these services, the important comparison is the complete workflow: preparation, execution, verification, revision and delivery.
Greater Autonomy Creates New Security Challenges
Astra’s capabilities also introduce substantial safety considerations.
In its September 3 safety report, OpenAI states that Astra is its first model to reach the Critical cybersecurity capability threshold under the company’s Preparedness Framework.
According to OpenAI’s evaluations, the model can identify previously unknown software vulnerabilities and develop sophisticated exploitation techniques when supplied with appropriate tools and access.
These capabilities can support legitimate defensive cybersecurity work, including identifying weaknesses and helping developers secure their software. They could also be misused.
OpenAI says it has therefore introduced stronger safeguards, including restrictions on advanced exploitation requests, additional monitoring and enhanced protections against attempts to bypass the model’s safety mechanisms.
There is another important distinction: a model that is better at completing tasks is not automatically easier to supervise.
OpenAI acknowledges that Astra’s internal reasoning traces have become less reliable as a way of identifying certain problematic behaviors. The company has consequently increased its monitoring and invested in additional alignment techniques.
For businesses, the practical lesson is to combine model-level safeguards with explicit permissions and operational controls. An agent that can read an internal report does not necessarily need permission to modify the original document or send it to an external recipient.
Who Can Use GPT-6 Astra?
As of September 25, 2026, Astra is available through several OpenAI products, although access and usage limits depend on the subscription and workspace configuration.
ChatGPT Plus subscribers have limited Astra usage in ChatGPT Work and Codex. Eligible Pro, Business and Enterprise subscribers have broader access, subject to their plan and organizational permissions.
GPT-6 Pro, powered by Astra, is available in regular ChatGPT conversations on eligible Pro, Business and Enterprise plans.
Work is intended for longer assignments involving research, analysis and finished deliverables. Codex focuses on development workflows, including repository operations, testing and debugging.
Developers can also access Astra through the OpenAI API using the gpt-6-astra model identifier.
At the published standard API rates, Astra costs $10 per million input tokens and $50 per million output tokens. Separate pricing applies to caching and accelerated processing.
Subscription users should also distinguish between access to the model and their available usage. Astra can consume the shared Work and Codex allowance faster than GPT-5.6 Sol, depending on the size and complexity of the assignment.
OpenAI maintains the current conditions in its ChatGPT Work and Codex documentation.
GPT-6 Sol and Luna: Astra Is Already Becoming a Family
On September 22, OpenAI expanded the GPT-6 generation with two additional models: GPT-6 Sol and GPT-6 Luna.
These models incorporate advances introduced with Astra while targeting different balances of performance, speed and cost.
OpenAI positions Sol and Luna as more affordable options for workloads that do not always require the capabilities of its flagship model. Both have been introduced in ChatGPT Work, Codex and the API.
This creates an important consideration for organizations deploying AI at scale. Using the most capable model for every operation may not be economically efficient.
A demanding research assignment or complicated software migration might justify Astra, while routine extraction, classification or simpler programming tasks could be assigned to a less expensive model after appropriate testing.
The relevant measure is not the cost of a single request, but the cost of producing an acceptable result, including failed attempts, corrections and human review.
How Should Businesses Evaluate Astra?
The most useful starting point is a clearly defined task with measurable results, rather than an attempt to automate an entire department immediately.
For example, a team could ask Astra to prepare a recurring report from approved source documents. It could then compare the result with its existing process, examining factual accuracy, completeness, formatting, execution time and the effort required for verification.
A second trial could introduce the kind of complications that regularly occur in real work: conflicting figures, missing information, inaccessible files or changes to the original requirements.
These tests reveal whether the agent identifies problems and requests clarification when necessary, rather than producing an apparently complete result based on unsupported assumptions.
Businesses should also establish clear rules for accessing sensitive information, modifying shared files and performing external actions.
The objective is not maximum autonomy at any cost. It is reliable delegation, with the appropriate level of human control for each task.
Conclusion: A Different Way to Use ChatGPT
GPT-6 Astra represents an important development in OpenAI’s transition from conversational assistance toward AI systems capable of executing substantial professional work.
Its improvements in computer use, coding, context management and document production could reduce the effort required for complex digital tasks. The introduction of Sol and Luna also shows how OpenAI intends to make this generation of capabilities available across different performance and cost requirements.
However, the distinction between an impressive demonstration and a dependable business workflow remains essential. Astra’s published results provide reasons to investigate its capabilities, not guarantees that every delegated task will succeed.
The practical question for organizations is therefore changing. Instead of asking only how well an AI can answer a question, they increasingly need to establish which tasks it can complete reliably, which tools it should access and where human judgment remains necessary.
Sources
- OpenAI — GPT-6 Astra: A New Generation of Intelligence
- OpenAI — Safety Overview: GPT-6 Astra, September 3, 2026
- OpenAI Help Center — ChatGPT Work and Codex
- OpenAI Developer Community — GPT-6 Sol and Luna Announcement, September 22, 2026
The difference is potentially significant. Instead of asking an AI assistant to explain how to complete a task, users can increasingly delegate parts of the task itself. Through ChatGPT Work and Codex, Astra can combine reasoning with tools and computer access to carry out activities that previously required extensive manual intervention.
But greater autonomy also creates new questions. How reliably can the model execute lengthy assignments? How much supervision does it need? And do its improved capabilities justify the additional cost?
What Is GPT-6 Astra?
GPT-6 Astra is OpenAI’s flagship model in the GPT-6 family. According to the company’s official announcement, its development combines advances in pre-training, reinforcement learning and alignment.
Its principal areas of improvement include software engineering, computer use, web research, scientific reasoning and professional tasks requiring several consecutive actions.
Astra is designed to work through complicated assignments while retaining the original objective. For example, a user might ask it to research an industry, analyze several documents and turn the findings into a presentation using an existing company template.
The important development is not that AI can generate a presentation. Earlier models could already produce office documents. Astra is intended to handle more of the surrounding workflow, from collecting information and interacting with applications to incorporating additional instructions without losing track of the original assignment.
From Chatbot to Computer-Using Agent
One of Astra’s most consequential capabilities is computer use.
Traditional language models primarily interact with the world through text and specially designed tools. Computer-using agents can go further by navigating graphical interfaces, entering information, opening documents and interacting with applications.
OpenAI says Astra can perform activities such as updating customer records, filling in forms, organizing calendars, researching information and testing websites. It can also work with applications that do not provide a dedicated API.
Consider a hypothetical business workflow. A project manager needs to prepare a monthly report using information from a CRM, several spreadsheets and internal documents.
An AI assistant might previously have summarized the information after someone collected and uploaded it. With suitable access, an agent powered by Astra could potentially retrieve the relevant material, compare figures, identify missing information and prepare a report using an approved template.
That does not mean the report should be accepted without review. The agent’s permissions, source selection, calculations and final output still need appropriate checks.
The distinction between a model’s intelligence and its access to external systems is important. Our guide to the Model Context Protocol explains how AI applications connect to information and tools, and why those connections require carefully defined permissions.
What Do Astra’s Benchmarks Actually Show?
OpenAI has published extensive evaluations comparing Astra with GPT-5.6 Sol and several competing models.
Three results illustrate the changes in the areas most relevant to professional use.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| OSWorld 2.0 — Computer use | 72.6% | 65.7% |
| Terminal-Bench 4.0 — Coding and terminal tasks | 57.9% | 37.3% |
| AutomationBench — Professional automation | 41.4% | 18.1% |
Source: OpenAI’s published GPT-6 Astra evaluations. Results depend on the benchmark version, execution environment and testing settings.
The particularly large improvement on Terminal-Bench suggests that Astra can handle more demanding technical workflows than its predecessor under the reported conditions.
On OSWorld, OpenAI also reports that Astra completed evaluated tasks in substantially less simulated time. This matters because an autonomous agent’s usefulness depends not only on whether it eventually reaches the correct answer, but also on the time and resources required.
However, the figures should not be interpreted as universal success rates. A benchmark result does not establish that Astra will complete the same proportion of tasks in a particular company’s software environment.
Users also need to consider the surrounding agent system, available tools, reasoning settings and verification requirements. Independent testing on representative workloads remains essential.
For more on interpreting these figures, see our article Top of the AI Benchmarks: Is It the Best Model for You?.
Codex: Better Support for Long Programming Tasks
Software development is another major focus of Astra.
Beyond generating code, advanced coding agents need to understand existing projects, inspect dependencies, run tests, investigate failures and avoid introducing unrelated changes.
One particularly interesting improvement concerns long development sessions.
When a conversation or coding task exceeds a model’s context window, previous systems often compress earlier information into summaries. This process can lose important details, including failed debugging attempts or architectural decisions.
OpenAI has introduced an experimental context-management approach for Astra in Codex. The system can preserve notes across context windows and search earlier context when it needs to retrieve requirements, previous test results or other relevant information.
This could be valuable for demanding tasks such as migrating an application to a newer programming language, refactoring a large codebase or investigating a complicated software defect.
However, better context management does not eliminate the need for version control, automated tests and human review. A development agent should still demonstrate that its changes work and respect the project’s constraints.
Professional Documents: The Quality of the Finished Result Matters
OpenAI also emphasizes Astra’s ability to produce documents, spreadsheets and presentations that follow existing templates.
The challenge here is not simply generating attractive content. A professional deliverable must preserve the relevant facts, respect formatting requirements and remain usable when transferred into the organization’s normal workflow.
Astra is designed to maintain these requirements as the task evolves. Users can introduce new constraints, ask questions about intermediate results or change direction without necessarily restarting the assignment.
This positions ChatGPT Work more directly within everyday business activities, from financial analysis and project reporting to research and document preparation.
Competition in this area is developing quickly. Anthropic is also expanding its document-oriented workspace, as discussed in our analysis of Claude Docs and Slides.
For organizations evaluating these services, the important comparison is the complete workflow: preparation, execution, verification, revision and delivery.
Greater Autonomy Creates New Security Challenges
Astra’s capabilities also introduce substantial safety considerations.
In its September 3 safety report, OpenAI states that Astra is its first model to reach the Critical cybersecurity capability threshold under the company’s Preparedness Framework.
According to OpenAI’s evaluations, the model can identify previously unknown software vulnerabilities and develop sophisticated exploitation techniques when supplied with appropriate tools and access.
These capabilities can support legitimate defensive cybersecurity work, including identifying weaknesses and helping developers secure their software. They could also be misused.
OpenAI says it has therefore introduced stronger safeguards, including restrictions on advanced exploitation requests, additional monitoring and enhanced protections against attempts to bypass the model’s safety mechanisms.
There is another important distinction: a model that is better at completing tasks is not automatically easier to supervise.
OpenAI acknowledges that Astra’s internal reasoning traces have become less reliable as a way of identifying certain problematic behaviors. The company has consequently increased its monitoring and invested in additional alignment techniques.
For businesses, the practical lesson is to combine model-level safeguards with explicit permissions and operational controls. An agent that can read an internal report does not necessarily need permission to modify the original document or send it to an external recipient.
Who Can Use GPT-6 Astra?
As of September 25, 2026, Astra is available through several OpenAI products, although access and usage limits depend on the subscription and workspace configuration.
ChatGPT Plus subscribers have limited Astra usage in ChatGPT Work and Codex. Eligible Pro, Business and Enterprise subscribers have broader access, subject to their plan and organizational permissions.
GPT-6 Pro, powered by Astra, is available in regular ChatGPT conversations on eligible Pro, Business and Enterprise plans.
Work is intended for longer assignments involving research, analysis and finished deliverables. Codex focuses on development workflows, including repository operations, testing and debugging.
Developers can also access Astra through the OpenAI API using the gpt-6-astra model identifier.
At the published standard API rates, Astra costs $10 per million input tokens and $50 per million output tokens. Separate pricing applies to caching and accelerated processing.
Subscription users should also distinguish between access to the model and their available usage. Astra can consume the shared Work and Codex allowance faster than GPT-5.6 Sol, depending on the size and complexity of the assignment.
OpenAI maintains the current conditions in its ChatGPT Work and Codex documentation.
GPT-6 Sol and Luna: Astra Is Already Becoming a Family
On September 22, OpenAI expanded the GPT-6 generation with two additional models: GPT-6 Sol and GPT-6 Luna.
These models incorporate advances introduced with Astra while targeting different balances of performance, speed and cost.
OpenAI positions Sol and Luna as more affordable options for workloads that do not always require the capabilities of its flagship model. Both have been introduced in ChatGPT Work, Codex and the API.
This creates an important consideration for organizations deploying AI at scale. Using the most capable model for every operation may not be economically efficient.
A demanding research assignment or complicated software migration might justify Astra, while routine extraction, classification or simpler programming tasks could be assigned to a less expensive model after appropriate testing.
The relevant measure is not the cost of a single request, but the cost of producing an acceptable result, including failed attempts, corrections and human review.
How Should Businesses Evaluate Astra?
The most useful starting point is a clearly defined task with measurable results, rather than an attempt to automate an entire department immediately.
For example, a team could ask Astra to prepare a recurring report from approved source documents. It could then compare the result with its existing process, examining factual accuracy, completeness, formatting, execution time and the effort required for verification.
A second trial could introduce the kind of complications that regularly occur in real work: conflicting figures, missing information, inaccessible files or changes to the original requirements.
These tests reveal whether the agent identifies problems and requests clarification when necessary, rather than producing an apparently complete result based on unsupported assumptions.
Businesses should also establish clear rules for accessing sensitive information, modifying shared files and performing external actions.
The objective is not maximum autonomy at any cost. It is reliable delegation, with the appropriate level of human control for each task.
Conclusion: A Different Way to Use ChatGPT
GPT-6 Astra represents an important development in OpenAI’s transition from conversational assistance toward AI systems capable of executing substantial professional work.
Its improvements in computer use, coding, context management and document production could reduce the effort required for complex digital tasks. The introduction of Sol and Luna also shows how OpenAI intends to make this generation of capabilities available across different performance and cost requirements.
However, the distinction between an impressive demonstration and a dependable business workflow remains essential. Astra’s published results provide reasons to investigate its capabilities, not guarantees that every delegated task will succeed.
The practical question for organizations is therefore changing. Instead of asking only how well an AI can answer a question, they increasingly need to establish which tasks it can complete reliably, which tools it should access and where human judgment remains necessary.