Written by M Akshay | Bestructure

Published: September 8, 2026, 11:00 AM IST | Last Updated: September 8, 2026, 11:00 AM IST

OpenAI has just rolled out GPT-6 Astra, and going by everything the company has shared so far, this isn’t just another routine model update. Astra is being positioned as OpenAI’s most capable release to date, built specifically to handle longer, multiple step tasks rather than simply answering a single prompt and stopping there. From writing and debugging entire codebases to navigating browsers and working across spreadsheets and presentations, Astra is clearly aimed at doing actual work rather than just generating text.

What’s made this launch stand out even more is the conversation happening around it. OpenAI president Greg Brockman suggested the model might represent a meaningful step towards artificial general intelligence, while also acknowledging that people will likely disagree on what actually counts as AGI in the first place. At the same time, OpenAI has been unusually direct about the risks tied to Astra’s capabilities, particularly around cybersecurity. Let’s break down what GPT-6 Astra actually is, what it’s built to do, and why its own creator is treating this release with visible caution.

What Exactly Is GPT-6 Astra?

At its core, GPT-6 Astra is a frontier AI model designed to reason through complex, multiple stage tasks instead of producing a single isolated response. OpenAI says the model brings together improvements across pretraining, reinforcement learning, and alignment, with a particular focus on situations where an AI system needs to understand a broader goal, figure out the right next step, and keep working through several stages until the task is genuinely complete.

This is really what separates Astra from a typical chatbot experience. Ask a regular AI assistant to build a presentation from a folder of documents, and it will likely summarise the material and hand you a rough outline. Astra, on the other hand, is designed to actually go through the files, organise the information, build the presentation itself, and then revise it further based on follow up instructions, functioning closer to a digital coworker than a simple question and answer tool.

How Does GPT-6 Astra Actually Work?

The simplest way to think about Astra is as a system that pairs reasoning with the ability to take action. Rather than treating every request as one isolated question, it can break a larger goal down into smaller steps, work through each one, evaluate what happened, and then decide what to do next. This is essentially what the industry now refers to as agentic AI, where the model doesn’t just explain how to complete a task, it attempts to carry out parts of that task on its own.

Take a spreadsheet task as an example. Instead of just telling you which formula to use, Astra can inspect the actual data, spot patterns, run calculations, update the relevant cells, and prepare a finished file. The same logic applies to software development, where the model can read through an existing codebase, write new code, run tests, catch errors, and revise its own work based on what those tests reveal.

OpenAI has spread Astra’s capabilities across a genuinely broad set of real world use cases rather than narrowing it to one specialty, and coding sits right at the centre of that pitch. The company has already paired Astra with Devin, the AI coding agent built by Cognition, and reports suggest it delivered noticeably better results on internal testing benchmarks the moment it was plugged in, producing clearer test reports and making it easier to follow what changes were made and why. For a developer working inside a large, messy, years old codebase, that kind of contextual understanding matters just as much as the model’s ability to write fresh code from scratch, since most real engineering work involves reading and fixing existing logic rather than starting from a blank file.

Away from pure code, Astra is also being marketed as something that can sit inside everyday office work. OpenAI describes it as capable of building and editing documents, spreadsheets, and slide decks directly from instructions and existing templates, effectively finishing a task end to end rather than handing back a rough draft for someone else to clean up. Brockman himself demonstrated this angle publicly, showing the model moving through spreadsheets and filling out forms at a pace he described as considerably faster than a person could manage manually, all while narrating the task out loud rather than typing individual commands.

That same computer use ability extends naturally into web based tasks too. Instead of simply returning a list of links after a search query, Astra can click through pages, gather the information it actually needs, and carry out an action based on what it finds, essentially behaving the way a person would while multitasking across browser tabs. Independent benchmark results back this up to some extent, with Astra scoring dramatically higher than OpenAI’s previous flagship model on ARC-AGI-3, a test specifically designed to check how well an AI system handles entirely unfamiliar scenarios rather than ones it may have already seen during training.

On the research and professional side, OpenAI is positioning Astra for scientific and knowledge heavy work as well, citing strong internal results across mathematics, cybersecurity, and general professional tasks. Some of that cybersecurity strength is being framed defensively too, with the company noting that security teams can use Astra for tasks like reviewing code for hidden flaws and applying patches before those flaws get exploited elsewhere. Even so, OpenAI has been careful to frame all of this as claimed performance rather than a promise that the model will handle every complex, real world task correctly without a human keeping an eye on it.

Why Is OpenAI Being So Cautious About Astra’s Cybersecurity Capabilities?

This is arguably the most notable part of the entire launch. It isn’t about what Astra can do with a spreadsheet or a block of code, it’s about what the model is capable of doing to computer systems themselves.

OpenAI has confirmed that Astra became the first model in its lineup to cross what it calls the Critical threshold under its own Preparedness Framework, an internal system built to assess whether increasingly capable AI models could enable serious new forms of harm. Reaching this Critical level essentially signals that OpenAI believes Astra’s capabilities go beyond simply making existing attacks easier, potentially enabling genuinely new categories of risk.

Part of the reasoning behind this comes from Astra’s ability to discover and exploit software vulnerabilities. According to OpenAI, the model achieved a perfect score on ExploitBench, a benchmark specifically designed to measure how effectively AI systems can exploit real world software flaws, comfortably ahead of the company’s previous flagship model. In a separate internal evaluation, OpenAI also says Astra managed to identify and successfully exploit two previously unknown vulnerabilities on its own.

To manage this risk, OpenAI says it has added stronger safeguards around Astra, including systems built to detect misuse, resist jailbreak attempts, and apply extra restrictions on higher risk users or use cases. The company hasn’t gone into full public detail about exactly how each of these restrictions functions though, which leaves an important tension at the centre of this entire release. The very same capabilities that make Astra genuinely useful as an autonomous agent are also what make it harder to fully control once it’s given real access to computers, software, and the internet.

GPT-6 Astra Pricing and Availability

OpenAI is rolling Astra out in stages rather than releasing it to everyone at once. Access initially went to a limited set of organisations, particularly those enrolled in OpenAI’s cybersecurity focused program, before gradually expanding to ChatGPT Plus, Pro, Business, and Enterprise subscribers. Developers can also access the model through the OpenAI API, Microsoft Azure, and Amazon Web Services.

Access Tier	Availability
Select organisations and cybersecurity program partners	Available first
ChatGPT Pro, Enterprise, Business Premium	Available in ChatGPT Work and Codex
ChatGPT Plus and standard Business users	Rolling out within a few days
API, Azure, AWS Bedrock	Available at launch

For API users specifically, GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens, working out to roughly Rs 940 and Rs 4,700 respectively at current exchange rates. Cached input costs $1 per million tokens, while cache writes are priced at $12.50 per million tokens. OpenAI also offers a faster processing mode for developers who need quicker responses, though it costs twice as much as the standard rate. Actual costs for users in India may vary slightly depending on exchange rates, applicable taxes, and individual billing arrangements.

For regular ChatGPT users, it’s worth noting that API token pricing doesn’t directly translate into subscription costs. Astra usage is included within existing Plus, Pro, Business, and Enterprise subscription limits, meaning subscribers don’t need to purchase a separate plan just to access the model, though additional usage credits can be purchased if needed. Free ChatGPT users currently continue using earlier models, and OpenAI hasn’t confirmed whether or when Astra might become available to that tier.

Also, read

Nvidia Acquires Hugging Face for 13 Billion Dollars: Here’s Why This Deal Matters

Is GPT-6 Astra Really a Step Towards AGI?

This question has understandably generated plenty of discussion since launch. Brockman’s comments suggesting the industry may have reached a meaningful AGI milestone drew reactions from across the tech world, including public commentary from other prominent industry figures who appeared to agree with that framing. At the same time, several researchers and commentators have pushed back, pointing out that there still isn’t a single, universally accepted definition of AGI in the first place, making any such claim somewhat subjective rather than a settled fact.

For everyday users and businesses, the more practical takeaway is simpler. Astra represents a genuine jump in what AI models can autonomously accomplish across coding, computer use, and professional workflows, regardless of whether that ultimately gets labelled as AGI or simply as a very capable step forward.

What This Means for Everyday Users and Businesses

For developers, Astra could meaningfully speed up work on large, complex codebases, particularly tasks that involve understanding existing logic and repeatedly testing changes rather than only writing fresh code. For professionals handling documents, spreadsheets, and presentations, it offers a shift from basic content generation towards actually completing full pieces of office work. And for businesses concerned about cybersecurity, Astra’s dual nature as both a powerful defensive tool and a potential risk if misused makes it a model worth adopting carefully, with proper oversight and controls in place rather than blind reliance.

Frequently Asked Questions

Q1. What makes GPT-6 Astra different from OpenAI’s earlier models? It’s built to handle longer, multiple step tasks involving reasoning and real action, including coding, document work, and browser based tasks, rather than just answering single prompts.

Q2. Why did OpenAI flag Astra under its Critical risk category? The model crossed OpenAI’s Critical threshold for cybersecurity capability after demonstrating the ability to discover and exploit software vulnerabilities, including previously unknown ones, during internal testing.

Q3. Who gets access to GPT-6 Astra first? Select organisations and cybersecurity program partners received early access, followed by ChatGPT Pro, Enterprise, and Business Premium users, with Plus and standard Business users following within days.

Q4. How much does GPT-6 Astra cost for developers? Through the API, it costs $10 per million input tokens and $50 per million output tokens, roughly translating to about Rs 940 and Rs 4,700 respectively at current exchange rates.

Q5. Does using Astra require a separate subscription for ChatGPT users? No, Astra usage is included within existing Plus, Pro, Business, and Enterprise subscription limits, though users can purchase additional credits if they exceed their allotted usage.

Q6. Is GPT-6 Astra considered true artificial general intelligence? That remains genuinely debated. OpenAI’s leadership has suggested it may represent a meaningful milestone, while several independent researchers argue there’s still no universally agreed definition of AGI to measure against.

Disclaimer: The details covered in this article have been gathered through research across OpenAI’s own official announcements along with multiple credible technology publications covering the GPT-6 Astra rollout. Every effort has been made to present this information accurately and in our own words, though readers are encouraged to check OpenAI’s official channels for the most current pricing, availability, and safety related updates.

Written by M Akshay for Bestructure.