OpenAI has released its brand-new flagship model – GPT-6 Astra, which the organization calls the most intelligent and aligned ever created. Astra is a significant innovation on a wide front, spanning from standard computer applications and internet searches to software development, cybersecurity, science discovery, and expert-level knowledge work, and OpenAI believes that this is more than just an evolutionary step.
The model will be rolled out in stages. It is now available to a limited number of companies and will be accessible to ChatGPT Plus, Pro, Business, and Enterprise users in the coming days. The model will also be available via the OpenAI API, Microsoft Azure, and AWS Bedrock for developers.
Record-Setting Benchmark Scores
Astro’s headline numbers are really amazing. OpenAI says the algorithm achieved 97.6% on FrontierMath Tier 4, that’s basically the top level of a series of extremely challenging mathematical benchmarks which include some of the world’s really tough open mathematical problems. Astra’s achievement is even more remarkable considering OpenAI has said that it already helped solve open mathematical issues. They also found that it was 99. 9% saturated on ARC-AGI-3 an evaluation for generating problem-solving strategies and abstract reasoning, and that it got all 100 on ExploitBench, a benchmark used to measure how well models can develop cybersecurity exploits.
A Major Focus on Alignment
OpenAI is positioning Astra as its most aligned model so far, meaning it better understands user intent and is less likely to act outside the scope of what it was asked to do. To test this, the company built a new evaluation inspired by an incident involving Hugging Face, checking whether a model given a difficult or impossible task would exceed its intended boundaries. OpenAI reports that GPT-5.6 Sol, Astra’s predecessor, did this 48% of the time when run without production safeguards. Astra did it 0% of the time.
The Best Computer-Use Model OpenAI Has Built
Computer usage is a major theme in the Astra release. Astra is said to manage tedious digital chores like filling-out an online form, updating CRM records, and sorting through calendars, with handling more complex tasks like doing research writing data science of a research paper, generating plots, building websites, and conducting frontend quality assurance. These improvements directly result in real gains in productivity.
On an OSWorld 2. 0 simulation, OpenAI said Astra computer tasks take 47% less time than GPT-5. 6, while at the same time, she scores higher: 72. 6% in roughly 40 minutes per task, against 65. 7% in approximately 75 minutes for GPT-5. 6 Sol, they mention. And, besides launchingAstra, OpenAI also has released an updated Codex harness that they claim reduces the Mind2Web benchmark task completion time by 1. 9 times compared to that of GPT-5. 6 Sol. Silas Alberti, SVP of Research at Cognition, mentioned thathis team is going to begin calling Astra through Devin’s harness on the day of the release.
Built for Professional Work
Besides mere computer use, OpenAI states that Astra learns how workspaces are organized and the kinds of things needed there. Astra works within current presentation and document templates, creates clean slides and documents, and cooperates with an existing company style in copying at a spreadsheet and producing financial and logistical analyses. Astra is also trained to include only pertinent context in its answers, to minimize revising.
And, it demonstrates improved visual judgment for website, game, and rendering creation. Astra can build and host websites, apps, and games using ChatGPT’s Sites feature all in one entry point. As Alex Mashrabov, CEO and co-founder of Higgsfield AI, commented, “Astra is a real edge for our company’s capability and efficiency.
Astra pulls off creative complex workflows that are required for a designer, while use of 20% less tokens of other models we have tested. OpenAI also notes Astra’s enhanced self-awareness which allows the model to better interpret input commands that are unclear. If it doesn’t have enough information to continue, Astra fills in common placeholders, avoiding unnecessary hestitation, only posing specific questions if the answer would alter the program’s result.
Astra is also capable of asynchronously polling the user with questions while completing other variables of the prompt; based on OpenAI. Astra also performs better at tracking a task, adding further instructions in a natural way, all while remaining true to initial commands of earlier models.
Stronger Coding Performance
OpenAI has labeled Astra as their most advanced model yet for programming. Based on John Crezepzhi, AI Assistants team at Jane Street, Astra not only achieved excellent results on the company’s internal coding metrics, but also has a way of communicating that makes it even easier for the developer to understand, thereby resulting in the code needing much fewer revisions before being able to be deployed.
Fabian Hedin, a co-founder and CTO of Lovable, mentioned that his group compared Astra in the three categories: low, medium, and high effort and found that it greatly outperformed GPT-5. 6 SOL even at the highest difficulty level where a larger number of iterations, in-depth browser-based validation, and a more aggressive focus on direct code execution are involved.
Apart from launching Astra, OpenAI has also added a new memorization module for Codex that enables Astra to record and retrieve the contextual flow as the context window gets filled up. Unlike a single compressed summary that gets created during extended sessions which tends to lose the explanation of why a fix failed or how a feature operates, Astra leaves all the older context searchable as it keeps notes across context windows. For now, this tool can be opted in via one line in Codex’s setup file while the change should soon become the new default behavior of Astra.
Advances in Science and Cybersecurity
At OpenAI, Astra is considered a significant advancement in scientific research, and its developers are even revealing two new findings related to the distribution of prime numbers as examples of Astra’s mathematical capability. Besides, Astra has set new records on various science and health benchmarks. OpenAI further notes that with its integration of scientific reasoning and computer use, Astra can be applied directly as a tool within specialized research software.
When it comes to cybersecurity, OpenAI honestly lays out the trade-offs. It states that Astra is a major improvement in cyber capabilities and the company even considers that it now satisfies the “Critical” threshold of its Preparedness Structure. While its ability to detect and develop zero-day exploits is an asset to defenders enabling patching of vulnerabilities at a faster rate, OpenAI admits this very feature increases the necessity of stronger safeguards.
In unrestricted testing, Astra achieved a 100% score on ExploitBench, while GPT-5. 6 Sol only got 78. 5%. On ExploitGym, Astra reached a 42. 4% success rate, against 30. 3% for its predecessor. Besides, during testing on a benchmark composed from a set of vulnerabilities that had been disclosed in the months before Astra’s release, Astra allegedly detected two previously undisclosed zero-day vulnerabilities of its own, which OpenAI says it will now report to software maintainers affected.
With these functionalities, the publicly available version of Astra is subject to limitations: it will support tasks like secure code review and patching but should deny more complex requests, like generating proof-of-concept exploits, at all times. OpenAI mentions that it is going to slowly remove these limitations for certified defenders via its Daybreak program in these weeks.
Deployment and Pricing
GPT-6 Astra is being offered by the OpenAI API under the model name gpt-6-astra, Microsoft Azure and Amazon Bedrock besides being directly by OpenAI. The price of the standard API, based on which you pay for input ($10), and output ($50) million tokens with the different rate for cache reads and writes. A faster processing mode is also being provided alongside that doubles both the speed and the standard price. Existing ChatGPT subscription plans cover the usage of the service. Users and businesses who want to exceed the available free usage can easily do so by purchasing additional credits.
Pro subscribers, Business subscribers and Enterprise subscribers will also have access to a version known as GPT-6 Astra Pro. Enterprise admins can enable Astra functionality to their workplace even if it is turned off by default upon release. Model supports Zero Data Retention for for eligible API customers, and besides that, OpenAI says they are developing a feature named Private Safety Processing which is intended to strengthen safety monitoring while keeping customer privacy intact.
FAQs
Q 1: What is GPT-6 Astra?
Open AI’s newest flagship model, designed to be its most capable and aligned model yet, with significant improvements in computer use, coding, science, and cybersecurity.
Q 2: When is it available?
It is available today to a limited set of organizations, and will roll out to ChatGPT Plus, Pro, Business, and Enterprise users over the next few days.
Q 3: What is its API pricing?
Standard price tier is $10 per million input tokens and $50 per million output tokens, with a faster but more expensive tier priced at twice the standard price.
Q 4: How is it safer/aligned?
Astra is much less likely to be “overstepping,” according to an internal benchmark. In a closed task, it exceeded a given target 0% of the time, compared to 48% for its predecessor, GPT-5.6 Sol.
Q 5: How is it better at computer-use tasks?
It is capable of completing tasks “about 47% faster than GPT-5.6 Sol,” while also achieving higher accuracy, according to OpenAI’s internal benchmarks on the OSWorld 2.0 suite of tasks.
Q 6: Is it good at cybersecurity tasks?
To an extent, according to OpenAI’s benchmarks. Astra achieves the “Critical” level required to pass the Preparedness Framework for cyber capabilities according to OpenAI’s standards— but due to its high level of preparedness, it declines to attempt several advanced tasks, including creating proofs of concept for cyber-attacks.
Q 7: Can it help with scientific research?
Astra has made gains in realms of scientific discovery, such as providing new insight into the distribution of prime numbers, and setting records on various science and health benchmarks.
Q 8: Is it good for coding?
“Astra represents our best model for software engineering to date,” the company said, and has seen endorsement from partners such as Jane Street and Lovable.
