ChatGPT’s New GPT-6 Astra: What Changed and Why It Matters

OpenAI has started rolling out its next-generation model, GPT-6 Astra, and this update looks different from the usual “better benchmark scores” release.

In ChatGPT, eligible users will see Astra through GPT-6 Pro, while the underlying model is also available as GPT-6 Astra through the API. OpenAI describes it as its most capable model so far, with major improvements in computer use, browsing, software engineering, cybersecurity, science and professional work. The rollout begins with selected enterprise customers, with Plus, Pro, Business and Enterprise access following over the coming days. (OpenAI Developers)

What makes Astra interesting is not simply that it can answer harder questions. The bigger change is that ChatGPT is becoming much better at doing multi-step work across different tools and applications.

Here are the features that matter most.

1. GPT-6 Astra Is Built for End-to-End Work, Not Just Better Answers

Older generations of large language models were mainly judged by how well they could answer a prompt.

Astra is being positioned differently.

OpenAI says GPT-6 Astra is designed for difficult end-to-end work involving complex reasoning, coding, research, computer use and document creation. It is particularly strong at carrying out workflows that move across code, browsers and professional software rather than staying inside a single chat response. (OpenAI Developers)

That distinction matters.

Consider a task such as researching a market, finding relevant sources, comparing companies, organizing the results, creating a spreadsheet and turning the findings into a presentation.

A chatbot can help with individual parts of that process.

A capable agent needs to understand the overall goal, decide what to do next, use the appropriate tools, keep track of intermediate results and continue until the work is finished.

Astra appears to be designed much more explicitly for the second type of work.

This is probably the most important shift in the new model.

ChatGPT is moving from answering questions toward completing workflows.

2. Computer Use Has Become a Core Capability

One of Astra’s biggest improvements is computer use.

OpenAI says the model reaches state-of-the-art performance on benchmarks designed to evaluate computer workflows across different professions. It is intended to operate browsers and software interfaces more effectively, rather than relying only on APIs or text-based tools. (OpenAI Developer Community)

This could have a much larger practical impact than another improvement in general reasoning.

A huge amount of office work still happens through graphical interfaces.

Employees log into websites, copy information between systems, update dashboards, upload documents, search internal portals and complete forms. Many of those systems were never designed with AI APIs in mind.

If an AI can reliably interact with those interfaces directly, the range of work it can automate expands dramatically.

The future enterprise AI agent may not need every application to expose a perfectly designed API.

It may simply use the software much like a person does.

That brings ChatGPT closer to something resembling a digital worker rather than a conversational assistant.

3. You Can Change Instructions While Astra Is Still Working

Another useful change is something OpenAI calls mid-turn steering.

With Astra, users can provide additional instructions while the model is already working.

Imagine asking ChatGPT to analyze 20 competitors and prepare a report. Halfway through, you realize you only care about companies in Asia.

Previously, long-running AI tasks often forced users to wait until the task finished, correct the instruction and start again.

Astra is designed to incorporate the new requirement while continuing the task. (OpenAI Developers)

That sounds like a relatively small feature, but it changes the interaction model.

Real work rarely follows a perfectly written prompt.

Requirements change.

New information appears.

Someone asks for a different format.

A manager suddenly wants another comparison included.

The ability to steer an AI while it is working makes the relationship feel less like:

Prompt → Wait → Result

and more like:

Assign → Observe → Adjust → Continue

That is much closer to how people actually collaborate with coworkers.

4. Astra Can Keep Working While Tools Are Running

GPT-6 Astra also introduces asynchronous tool calling for developers.

Traditionally, when an AI calls an external tool, it often has to wait for that tool to return a result before continuing.

Astra can continue reasoning, call other tools or work on independent parts of the task while another operation is still running. (OpenAI Developers)

This is particularly important for agentic AI.

Imagine an AI conducting due diligence on a company.

It might simultaneously:

search public filings,

query an internal database,

analyze financial documents,

check recent news,

and prepare a comparison table.

Some of those operations may take several seconds or even minutes.

If the agent has to perform them one by one, complex workflows become painfully slow.

Async tool calling makes parallel work possible.

The model becomes less like someone following a checklist sequentially and more like a project manager coordinating several tasks at once.

For developers building AI agents, this may end up being one of Astra’s most important architectural improvements.

5. The Context Window Is Now Large Enough for Very Big Projects

GPT-6 Astra supports a 1,050,000-token context window and up to 128,000 output tokens through the API. Its published knowledge cutoff is April 30, 2026. (OpenAI Developers)

A million-token context window changes what kinds of projects can realistically fit into a single working context.

Instead of providing a few documents, developers can potentially give the model:

large code repositories,

long research collections,

contract libraries,

technical documentation,

multiple financial reports,

or extensive project histories.

The important point is not simply that ChatGPT can “remember more.”

Large context is valuable because complex work often depends on relationships between information scattered across many files.

A model reviewing a software repository may need to understand how code in one directory affects a function hundreds of files away.

A model reviewing legal documents may need to compare terms across dozens of agreements.

A smaller context window forces developers to retrieve and feed information piece by piece. A much larger one gives the model a broader view of the problem.

That can make AI agents less fragmented when handling large professional tasks.

6. Reasoning Is Becoming Something Users Can Control

OpenAI has also been moving away from the idea that every question should receive the same amount of reasoning.

GPT-6 Astra supports several reasoning levels through the API, from low through medium, high, xhigh and max. (OpenAI Developers)

In ChatGPT, this fits into the broader shift already introduced with GPT-5.6, where users can choose how much thinking they want the model to spend on a task.

The logic is straightforward.

You do not need maximum reasoning to rewrite an email.

You may want it when debugging a complicated system, analyzing a contract or working through a difficult research problem.

This creates a more practical trade-off between:

speed, cost and intelligence.

The best AI experience may no longer come from always selecting the smartest possible mode.

It may come from using exactly as much intelligence as the task requires.

7. Coding Is Becoming More Agentic Too

Software engineering remains one of the areas where OpenAI says Astra has made significant gains.

The difference, again, is not limited to writing individual functions.

Astra is intended for multi-step software engineering workflows, including navigating repositories, running tools, making changes and dealing with larger projects. OpenAI describes it as state-of-the-art across software engineering and terminal-based tasks. (OpenAI Developer Community)

This fits a larger shift happening across AI coding tools.

The first generation of coding assistants mostly completed code.

The next generation answered programming questions and generated larger blocks.

Coding agents now increasingly take responsibility for entire issues:

read the repository,

find the relevant files,

modify the code,

run tests,

debug failures,

and prepare the final result.

Astra appears designed for exactly this kind of workflow.

For developers, the question is becoming less:

“How good is the model at writing code?”

and more:

“How much of the software engineering process can the agent reliably complete?”

8. Cybersecurity Capability Has Reached a New Level

One of the most striking parts of the Astra release has little to do with productivity.

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

That is a significant threshold.

According to OpenAI, with the right tools and access, Astra may be capable of finding previously unknown vulnerabilities and developing exploitation methods across well-protected systems without requiring a human to guide every individual step. (OpenAI)

That capability is obviously useful for defensive cybersecurity.

Security teams could potentially use models like Astra to discover weaknesses before attackers do.

But the same capability creates obvious risks.

OpenAI says it has therefore introduced stricter isolation, stronger model protections, extensive trajectory monitoring, automated red-teaming and additional safeguards around high-risk cyber activity. (OpenAI)

This part of the release is worth watching closely.

For years, discussions about powerful AI models were often hypothetical.

A model reaching a formal “Critical” capability threshold shows that the safety problem is becoming much more concrete.

9. Astra Is Also Designed to Follow Intent More Carefully

Capability is only useful if the model understands what the user actually wants.

OpenAI describes Astra as its most aligned model so far, with improvements in understanding intent, respecting task boundaries and responding to changing instructions. (OpenAI)

One practical improvement is how it handles ambiguity.

Instead of asking unnecessary questions about routine details, Astra is designed to infer reasonable defaults from context.

But when an ambiguous instruction could materially change the result, it is supposed to ask a focused question.

That balance is surprisingly important.

AI agents become frustrating if they constantly stop and ask for confirmation.

But they become dangerous if they guess too aggressively.

A useful agent needs to know when it can proceed and when it should stop.

That sounds less impressive than a benchmark score, but for everyday work it may be one of the most important improvements.

10. GPT-6 Astra Is Powerful, but It Is Not Replacing Everything

Another important detail is that Astra does not replace the entire GPT-5.6 family.

GPT-5.6 Sol remains the flagship option for many complex tasks, while Terra and Luna provide lower-cost alternatives. Astra sits above them for the hardest workflows. (OpenAI Help Center)

OpenAI’s API pricing illustrates the difference.

GPT-6 Astra currently costs $10 per million input tokens and $50 per million output tokens, compared with $4 and $20 respectively for GPT-5.6 Sol. (OpenAI Developers)

So Astra is substantially more expensive per token.

However, OpenAI argues that the model can complete some difficult tasks using fewer output tokens and therefore achieve a lower total cost per successful task despite the higher token price. (OpenAI Developers)

This reflects a broader shift in AI economics.

The cheapest model is not always the cheapest way to finish a job.

If one model requires five attempts and another succeeds on the first, token price alone tells only part of the story.

For businesses, the more meaningful metric may increasingly become:

cost per completed workflow.

Who Can Use GPT-6 Astra?

GPT-6 Astra is being rolled out gradually.

OpenAI says access begins with selected enterprise organizations, with the API and ChatGPT Plus, Pro, Business and Enterprise availability following during the rollout. In ChatGPT, eligible users will access Astra as GPT-6 Pro. (OpenAI Developers)

Availability may differ between Chat, ChatGPT Work and Codex while the rollout is underway.

Enterprise administrators also have additional model-access controls, and Astra may initially be disabled by default in some managed workspaces. (OpenAI Help Center)

So if GPT-6 Pro does not appear immediately in your model picker, that does not necessarily mean your plan will not receive it.

The rollout is still in progress.

The Bigger Change: ChatGPT Is Becoming Less Like a Chatbot

The easiest way to describe GPT-6 Astra would be to say that it is smarter than GPT-5.6.

That is probably true, but it misses the more interesting change.

The most important improvements are happening around action.

Astra can work across software, coordinate tools, continue while external operations are running, accept new instructions mid-task and handle much larger working contexts.

Those capabilities move ChatGPT further away from the original question-and-answer interface.

The trajectory is becoming clearer:

Chatbot → Copilot → Agent → Digital Worker

That does not mean autonomous AI is suddenly ready to run every business process. Reliability, security, permissions and human oversight remain major constraints.

But GPT-6 Astra gives a clearer picture of what OpenAI thinks the next version of ChatGPT should become.

Not simply something you ask.

Something you give work to.

And that may ultimately be the most important feature of the new ChatGPT. (OpenAI Developers)