On September 3, 2026, OpenAI unveiled GPT-6 Astra. The model navigates software on its own, fills out forms, and works on tasks that span hours or days. For controlling teams that are just beginning to establish approval processes for simpler AI agents, this raises a new question: How much autonomy is acceptable in their own reporting and planning processes?
GPT-6 Astra is OpenAI’s new flagship model. Unlike earlier versions of ChatGPT, it doesn’t just act as an assistant that answers questions. It independently performs multi-step tasks on a computer, such as filling out forms, navigating web applications, or independently fixing programming errors.
What can GPT-6 Astra do that earlier models couldn’t?
GPT-6 Astra processes a context window of 1,050,000 tokens and generates up to 128,000 tokens of output in a single response. This is sufficient to process extensive reporting documents, contracts, or multi-year planning data in a single pass.
In software development, the model goes beyond mere code generation. It executes code, finds errors, and fixes them on its own. The game developer Playco reports a reduction of about 50 percent in manual corrections to prototypes compared to the previous model.
In research, an internal version of Astra solved ten problems in mathematics and computer science theory, some of which had remained unsolved for more than ten years. People provided the initial lines of reasoning, and the model then formalized them into machine-verifiable proofs. According to OpenAI, the tokens required for this cost around $2,000 – a fraction of the six- to seven-figure budgets that research teams typically allocate for comparable problems.
Why does OpenAI classify Astra as a critical-level security risk?
OpenAI evaluates new models internally using a security framework that classifies capabilities into categories such as cybersecurity, biology, and chemistry. According to OpenAI, Astra is the first model to exceed the critical threshold in the cybersecurity category.
This has direct implications for everyday life. In ChatGPT and Codex, the system may require additional confirmation before performing certain actions. Through the API, individual tasks can be interrupted if they are classified as potentially risky.
For controlling teams, this means, in practical terms: An agent that has access to ERP systems, planning tools, or payment approvals has different requirements for authorization models than a chatbot that only responds with text. Anyone who already works with AI agents is somewhat familiar with this difference. We explain where the line between an assistant and an agent lies in our article on AI agents in controlling.
How does Astra fit into your approval processes?
We’ll work with you to figure out how autonomous AI agents fit in – even if the outcome is that you’re not quite ready for that yet.
What steps make sense for controlling teams?
- Review existing approval processes: Where do humans make decisions today, and where could an agent make a recommendation in the future?
- Document access rights to financial systems before integrating an agent.
- Choose a pilot project with a limited scope, such as data preparation rather than payment approval.
- Randomly cross-check the agent’s results, especially during the first few weeks.
- Involve governance stakeholders before connecting to production systems.
What This Means for BI2run Customers
New models like GPT-6 Astra show the direction in which AI agents are evolving. For controlling teams, this doesn’t mean implementing every new model immediately. It means setting up their own system architecture so that it can keep pace with this evolution when the time is right.
BI2run supports controlling and finance teams with this very process – from the initial evaluation of a new model to its integration with Planning Analytics or Power BI. If an implementation isn’t feasible at this time, we’ll say so.
- Launch: September 3, 2026, OpenAI
- Rollout: initially enterprise customers in the Daybreak program, then ChatGPT Plus, Pro, Business, Enterprise, followed by API and AWS
- Context window: 1,050,000 tokens, output up to 128,000 tokens
- Knowledge cutoff: April 30, 2026
- Pricing: $10 per 1 million input tokens, $50 per 1 million output tokens
- First OpenAI model with a critical rating in the cybersecurity category
- US comparison figure: search volume for “ChatGPT Agent” grew from 320 to 10,800 queries per month within twelve months
Source: OpenAI announcement dated September 3, 2026, SE Ranking keyword database (US), as of September 2026.
Glossary
Frequently Asked Questions About GPT-6 Astra
What is the difference between GPT-6 Astra and ChatGPT Agent?
ChatGPT Agent was a separate product designed for browser-based tasks. Astra brings comparable capabilities directly into the base model and expands them to include programming and scientific work.
When will GPT-6 Astra be available for businesses?
Access will begin with Enterprise customers in the Daybreak program. ChatGPT Plus, Pro, Business, and Enterprise users will follow in the days after the launch, with API integration and AWS availability coming a little later.
Is GPT-6 Astra suitable for controlling processes?
The model can fill out forms, navigate systems, and process data in multiple stages. Whether this makes sense for a specific controlling setup depends on the existing system landscape and the approval processes.
What risks are associated with the use of autonomous AI agents such as Astra?
OpenAI itself classifies the model as critical in the area of cybersecurity and has therefore incorporated additional verification steps. Companies should adjust their own authorization policies accordingly before connecting agents to production systems.
How does GPT-6 Astra differ from a traditional AI assistant?
An assistant answers inquiries. An agent like Astra independently plans and carries out multiple steps, sometimes over the course of hours or days.
How much does it cost to use GPT-6 Astra via the API?
OpenAI charges $10 per 1 million input tokens and $50 per 1 million output tokens. This pricing is comparable to that of other current state-of-the-art models.
