GPT-6 Astra takes control of the computer under tighter oversight

OpenAI has launched GPT-6 Astra, a general-purpose model designed to use computers, write code, browse the web, and conduct scientific work, with cybersecurity capabilities classified at the Critical threshold.

Filling out a form, updating a customer record, searching for an apartment, fixing an application, or building a circuit in KiCad. With GPT-6 Astra, OpenAI no longer presents computer use as a specialized feature. The model is expected to observe an interface, choose its actions, check the results, and continue working for several dozen minutes.

GPT-6 Astra succeeds GPT-5.6 Sol at the top of OpenAI’s general-purpose model lineup. The company describes it as its strongest model for computer use, web browsing, coding, cybersecurity, science, and professional work.

The rollout begins with a limited number of organizations before gradually expanding to ChatGPT Plus, Pro, Business, and Enterprise subscribers. Pro, Business, and Enterprise accounts are also expected to receive GPT-6 Astra Pro, a version intended for the most demanding requests. Enterprise administrators must enable the model manually because access is turned off by default at launch.

The model is also coming to the API under the identifier `gpt-6-astra`, followed by Microsoft Azure and Amazon Bedrock. Standard pricing is set at $10 per million input tokens and $50 per million output tokens. Cache reads and writes use a separate pricing schedule.

Those rates are higher than GPT-5.6 Sol, which costs $5 per million input tokens and $30 per million output tokens. Astra therefore doubles the price of input and raises output costs by approximately 67%. OpenAI argues that part of the difference can be offset by shorter answers and fewer attempts on some tasks.

A Fast mode promises up to twice the speed at twice the standard price. If the multiplier is applied directly, that would amount to $20 per million input tokens and $100 per million output tokens. The company has not published a uniform latency guarantee across workloads, regions, or tool configurations.

Usage in ChatGPT is included within the existing limits of eligible subscriptions. Customers will be able to purchase additional credits after reaching those limits. No access for free accounts has been announced.

The main change comes from bringing reasoning and action together. Astra does more than produce instructions for a person to follow. It can operate a browser or application, interpret what appears on the screen, enter information, and adjust its route when the interface does not behave as expected.

The demonstrations cover a wide range of tasks: finding a pediatrician, booking a government appointment, comparing apartments, preparing a tax form, testing a website, working in Power BI, formatting a legal document, and creating a game.

In a more technical example, the model works directly in KiCad to turn an electronic schematic into a printed circuit board layout. It positions components and routes their connections inside the software, a process that requires understanding the project’s structure as well as operating the interface.

OpenAI also shows a house modeled in Blender and then imported into Unreal Engine 5 to create a walkable environment. The example demonstrates an ability to move between applications, but it does not measure architectural quality, building-code compliance, or project optimization.

In ChatGPT, the Sites feature is expected to use Astra to create and host websites, web applications, and games from a prompt. The model can write the code, start the project, inspect the result, and correct issues it encounters. This feedback loop moves generation closer to completed execution, although human review remains necessary before publishing anything sensitive.

OpenAI measures this ability on Agents’ Last Exam, a benchmark built around professional tasks carried out in real software. Astra scores 59.3%, compared with 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol in the company’s published comparison.

OpenAI says Astra uses approximately 65% fewer output tokens than Claude Opus 5 in the configurations that produced their highest scores. The comparison depends on reasoning settings, available tools, and the limits given to each model. Fewer tokens can reduce cost and waiting time without necessarily making every intermediate step easier to inspect.

On OSWorld 2.0, which evaluates the autonomous use of computer environments, Astra scores 72.6% while spending approximately 40 minutes per task. GPT-5.6 Sol reaches 65.7% in about 75 minutes. OpenAI calculates a 47% reduction in task time.

That result comes from a latency simulation rather than a universal measurement across all devices. Actual time will depend on the connection, applications being used, required confirmations, and number of actions performed.

OpenAI is also updating the system surrounding the model in Codex. The combination of Astra and this new setup reportedly completes Mind2Web tasks 1.9 times faster than the current GPT-5.6 Sol experience. The improvement cannot be attributed solely to the model because the execution environment changed at the same time.

Astra is also designed to adjust its behavior when a request leaves room for interpretation. OpenAI says it should make routine decisions independently, ask a question when the answer could materially change the result, and wait for approval before making consequential choices.

In Codex, those questions can be sent without stopping all other work. The model continues with steps that do not depend on the answer and incorporates the new information when it arrives. If no response is received, it can proceed with a reasonable assumption for minor decisions but should stop when an important choice is involved.

This distinction still depends on the model’s judgment. A decision that appears minor could affect a budget, modify a shared file, or disrupt a production service. Confirmation rules configured around the agent therefore remain essential.

Office work is another major focus of the release. Astra was trained to produce documents, presentations, and spreadsheets that follow an existing template. It is expected to reproduce the structure, tone, and visual