SecurityBrief US - Technology news for CISOs & cybersecurity decision-makers
United States
OpenAI launches GPT-6 Astra for paid ChatGPT users

OpenAI launches GPT-6 Astra for paid ChatGPT users

Thu, 3rd Sep 2026 (Today)
Sean Mitchell
SEAN MITCHELL Publisher

OpenAI has launched GPT-6 Astra, which it describes as its most advanced model to date.

The rollout is starting with a limited group of organisations before expanding to ChatGPT Plus, Pro, Business and Enterprise users, as well as to developers through the OpenAI API and AWS.

OpenAI presented Astra as a step forward in computer use, software engineering, scientific work, cybersecurity and other professional tasks. The model was built on advances in pre-training, reinforcement learning and alignment research.

Canva was among the early users given access ahead of the launch. Danny Wu, Head of AI Products at Canva, described how the model performed across the design software group's internal systems.

"Astra delivered some of the best and highest-signal recommendations we've seen. It effectively navigated our entire codebase of more than 80 million LoC, wrote and analysed more than 1,000 queries against our data warehouse, and used more than 21 different internal knowledge sources and community feedback to identify what delivers the most value," said Danny Wu, Head of AI Products at Canva.

Computer use

OpenAI said Astra can complete multi-step computer-based tasks, including producing documents, spreadsheets and presentations, as well as creating websites and testing whether features work. In latency simulations on the offline subset of OSWorld 2.0, the model achieved higher computer-use performance in about 47% less time per task than GPT-5.6 Sol, according to the company.

OpenAI also positioned Astra as a stronger tool for software engineering, saying it handled more complex work in real codebases and outperformed GPT-5.6 Sol on DeepSWE v1.1 at about 57% lower estimated API cost per task when comparing each model's highest-scoring setting.

Science focus

OpenAI said an internal version of Astra had already contributed to new mathematical results. It pointed to 10 advances in mathematics and theoretical computer science produced by that version of the model, with proofs formalised in Lean.

According to OpenAI, Astra scored 98% on FrontierMath Tier 4, a benchmark used to test advanced mathematical reasoning by frontier AI systems.

Alignment and oversight

OpenAI said Astra showed improvements in understanding user intent, respecting task boundaries and communicating clearly about its work. The model uses context to fill routine gaps and asks follow-up questions when the answer could change the result, the company said.

That emphasis on oversight reflects a broader commercial push in workplace AI tools, as companies try to make systems more useful while limiting the risk of incorrect or unauthorised actions. For business users, the pitch is not only better performance but greater confidence in assigning more complex work to a model.

Cybersecurity concerns

The launch also highlights the tension facing AI developers as models improve in sensitive areas such as cybersecurity. OpenAI said Astra's cyber capabilities could help defenders identify and fix weaknesses, while also acknowledging the need for tighter protections against misuse.

Under OpenAI's Preparedness Framework, Astra meets the Critical threshold in cybersecurity. The company said it is strengthening safeguards around the model and plans to expand access for work including vulnerability validation, malware analysis and detection engineering.

The release comes as leading AI companies compete to show practical gains in specialised business and technical work, rather than consumer-facing novelty alone. Benchmarks tied to coding, research, mathematical reasoning and computer-based task execution have become a central battleground as companies try to persuade customers that newer models can justify higher subscription and infrastructure spending.

For OpenAI, Astra's launch also underlines the role of external testing by large customers before broader deployment. Canva's account offers one example of how major software groups are assessing these systems against internal codebases, data warehouses and knowledge stores before deciding where they fit into daily workflows.

GPT-6 Astra will become available across OpenAI's paid ChatGPT tiers, as well as through its developer channels and AWS.