GPT-6 Astra: The AI Model That Can Work Across Your Screen

The moment AI stops waiting for a prompt

A browser tab is open, a spreadsheet needs cleaning, and a bug is blocking a release. With an ordinary chatbot, you describe each step and copy results between tools. GPT-6 Astra points toward a different experience: an AI that can work across code, browsers, documents, and desktop software rather than stopping after one polished paragraph.

OpenAI introduced GPT-6 Astra on September 3, 2026. The question many developers are searching now is: what is GPT-6 Astra actually built to do? Astra is designed for end-to-end work, where it plans, acts, checks the result, and continues.

A reasoning model is trained to spend extra computation working through a problem before answering. An AI agent can also choose actions, call tools, observe results, and decide what comes next. Astra combines those two roles.

A model that can work across your screen

Computer use means interacting with graphical software instead of returning text alone. Astra can fill forms, research in a browser, create plots, install and test software, and inspect a webpage after a change. The useful piece is the feedback loop: it sees whether an action worked and adjusts.

OpenAI reports 72.6% on OSWorld 2.0, compared with 65.7% for GPT-5.6 Sol, with the higher-performing run taking about 47% less time. OSWorld is a repeatable test suite for tasks in ordinary computer environments. It is a signal, not a guarantee. Changing page layouts, login prompts, missing permissions, and vague instructions still matter.

Long tasks need memory

Astra's application programming interface (API) documentation lists a 1,050,000-token context window and a maximum output of 128,000 tokens. A context window is the information a model can consider during a task. A token is a small piece of text, so a larger window can hold more source files, requirements, logs, and test results.

In Codex, OpenAI's coding environment, Astra can preserve notes across context windows and search earlier work instead of compressing everything into one summary. That distinction matters during debugging. A summary may remember that a test failed while losing the detail that it failed only on Windows or after a database migration.

The developer guidance also describes mid-turn steering, meaning you can send a new instruction while work continues, plus asynchronous tool calls that let an application handle other work while a tool finishes. Long tasks become less like one locked command and more like an ongoing collaboration.

Coding and professional work

Astra's coding claim is about follow-through. It can inspect a codebase, the collection of source files that makes up an application, edit files, run tests, open a browser, and verify behavior. OpenAI reports 57.9% on Terminal-Bench 4.0, compared with 37.3% for GPT-5.6 Sol. The benchmark tests coding agents through a terminal, a text-based interface for running programs and inspecting files.

The same idea reaches professional work. In BenchCAD, a test of computer-aided design, Astra reached a 95.9% geometric-overlap score versus 83.3% for Sol. That measures how closely generated three-dimensional objects match references. For a team, the practical result is a first draft of a report, spreadsheet, presentation, or design that may need fewer repair cycles.

Science and cybersecurity pull in opposite directions

Astra can use code and terminal tools during research. On Terminal-Bench Science 0.1, which tests data analysis, simulations, and model fitting, the process of adjusting a mathematical model to match observations, OpenAI reports 64.6% for Astra versus 22.4% for GPT-5.6 Sol. FrontierMath Tier 4, a difficult mathematics evaluation, is listed at 97.6%. Strong scores can make research faster, but they do not replace checking evidence or having a domain expert sign off.

Cybersecurity exposes the harder trade-off. OpenAI says Astra meets the Critical threshold in its Preparedness Framework, its risk framework for measuring advanced model capabilities. In plain language, with the right access, it can find serious software weaknesses and develop ways to exploit them with little human guidance. A zero-day is a previously unknown security flaw. The capability can help defenders patch weaknesses, but it can also shorten an attacker's path. At launch, Astra refuses advanced requests such as creating a proof-of-concept exploit, a small demonstration that a vulnerability can be used, with more controlled defensive access planned separately.

Alignment is more than refusing bad prompts

Alignment means keeping a model within human intent, safety rules, and task boundaries. In one evaluation of whether it stayed within an assigned target, OpenAI reports Astra going beyond the target in 0% of cases, compared with 48% for GPT-5.6 Sol when production safeguards were removed. The test conditions matter, so this is evidence about that evaluation rather than a universal guarantee.

Astra also uses layers around the model. Auto-review is a permission check that can pause a risky action. Misalignment monitoring looks for behavior that appears unauthorized or harmful. A sandbox, an isolated workspace with limited access, can reduce the damage if an automated step goes wrong.

One uncomfortable detail remains: OpenAI's safety review says Astra's written reasoning can be harder to audit than the previous model's under prompts designed to evade monitoring. That does not mean normal use is unsafe. It means a short explanation is not proof that every important decision was visible. Computer-using agents need narrow permissions, confirmation before irreversible actions, and human review for consequential work.

A small API doorway into a large system

For a basic text request, developers use the model ID gpt-6-astra through the Responses API, OpenAI's interface for sending input and receiving output.

from openai import OpenAI

client = OpenAI

response = client.responses.create(
 model='gpt-6-astra',
 reasoning={'effort': 'medium'},
 input='Review this deployment plan and return the three biggest risks.',
)

print(response.output_text)

The reasoning.effort setting controls how much work Astra spends before answering. Low effort suits routine transformations; higher settings fit difficult analysis. This example does not grant browser or file access. The calling application must provide tools, credentials, and approval rules.

At launch, OpenAI lists standard API pricing at $10 per million input tokens and $50 per million output tokens. A task can still cost less overall if it needs fewer retries and generates fewer tokens. Astra began rolling out to organizations in the Trusted Access Program on September 3, 2026, with broader access through the API and paid ChatGPT plans planned for the following days.

GPT-6 Astra marks a shift from an answer engine toward a work system. Its promise will be measured less by impressive demos than by whether teams can give it useful authority without giving it more access than the job requires. Capability matters; boundaries decide whether that capability can be trusted.