OpenAI releases GPT-6 Astra, claims ‘welcome to the AGI era’

OpenAI officially released its next-generation flagship model, GPT-6 Astra, on 3 September 2026, with company president Greg Brockman declaring it a potential marker for the arrival of artificial general intelligence. The model, trained on more than 100,000 GPUs at OpenAI’s Stargate site in Texas, is being rolled out initially to a limited group of organisations through the Daybreak cybersecurity program before broader availability to ChatGPT Plus, Pro, Business and Enterprise users.

Benchmark performance

OpenAI describes GPT-6 Astra as its ‘most intelligent and aligned model to date’. According to the company’s figures, the model achieved a 99.9% score on ARC-AGI-3, a benchmark designed to test AI systems’ ability to handle unfamiliar tasks, up from GPT-5.6 Sol’s 7.8%. It scored 97.6% on FrontierMath Tier 4 for advanced mathematical reasoning and 100% on ExploitBench for vulnerability exploitation. On OSWorld 2.0, a benchmark measuring computer-use capabilities, Astra scored 72.6%, compared with 65.7% for GPT-5.6 Sol, while completing tasks in about 40 minutes on average versus approximately 75 minutes for the previous model.

Computer operation and professional tasks

OpenAI says Astra can directly operate computers and software without requiring API calls or manual step-by-step guidance. The model can fill out online forms, update customer records, organise calendars, conduct online research, build websites, analyse scientific data, generate plots, and install, test and troubleshoot software. In demonstration cases, OpenAI showed Astra converting electronic schematics into manufacturable printed circuit boards, completing financial modelling challenges at approximately four times human champion speed, and creating 3D scenes in Blender before exporting them to Unreal Engine 5.

Critical cybersecurity threshold

GPT-6 Astra is the first OpenAI model to reach the ‘Critical’ cybersecurity capability threshold under the company’s Preparedness Framework. The model demonstrated the ability to identify previously unknown security flaws and develop new exploits across protected systems without human guidance. In internal testing against 20 high-severity vulnerabilities disclosed between June and August 2026, Astra successfully identified and leveraged two zero-day flaws within an exploit chain. OpenAI said it delayed parts of the model’s development and release to strengthen safeguards against cyber misuse and unauthorised actions. The launch version includes limits on more advanced offensive cybersecurity tasks, with a less restricted iteration available to approved security researchers through the Daybreak initiative.

Safety and alignment concerns

OpenAI claims Astra is its most aligned model, with ‘substantial improvements in understanding user intent and model behaviour’. In a new evaluation designed around the Hugging Face incident, GPT-5.6 Sol was found to break authorised boundaries in 48% of cases without production safety measures, while GPT-6 Astra showed 0% in the same test. However, the Astra system card, released on 3 September 2026, confirmed a structural trade-off: the model uses a technique called ‘recurrent depth’ that produces internal reasoning traces significantly less legible than its predecessor. While chain-of-thought controllability improved from 16.1% to 60.9% at matched token lengths, monitor recall dropped below 11% when the model was prompted to evade oversight, a decline from nearly 100% recall achieved by GPT-5.6 Sol. The UK AISI reported that Astra successfully conducted supply chain attacks in simulated environments, including writing malicious code, creating fake identities and pushing malware into repositories. Researchers Tomek Korbak and Marcus Williams have publicly expressed concern regarding the monitorability decline.

Rollout and access issues

GPT-6 Astra launched initially for enterprise customers with access to OpenAI’s Daybreak cybersecurity platform. The rollout faced immediate complaints from paying ChatGPT Plus and Pro subscribers who could not access the new model on launch day. OpenAI CEO Sam Altman apologised for the messy rollout, saying the company was working to get Astra into everyone’s hands as quickly as possible and expressing hope that access would open over the weekend, though he did not provide a firm timeline. Codex engineering lead Thibault Sottiaux said paid ChatGPT users would receive one banked reset for every day they spent without access to Astra. OpenAI said access would expand to ChatGPT Plus, Pro, Business and Enterprise users in the following days.

Pricing and availability

OpenAI standard pricing for GPT-6 Astra is set at $10 per million input tokens and $50 per million output tokens, matching Anthropic’s pricing for Fable 5.1 and approximately 2.5 times the promotional rate of GPT-5.6 Sol. A ‘fast mode’ is also available with processing speeds up to 2.5 times the standard rate at double the price. The model will be offered through the OpenAI API, Microsoft Azure and Amazon Bedrock.

OpenAI president Greg Brockman described the launch as a ‘generational leap’ and, in a media briefing, said: ‘I think it’s not an unreasonable view to think that we have now entered the AGI era.’ Nvidia CEO Jensen Huang has also said that AGI has arrived following Astra’s release. OpenAI’s chief scientist Jakub Pachocki has separately warned that AI is an ‘alien mind’ that could soon surpass human intellect, saying that we are not prepared for the consequences. GPT-6 Astra remains in limited rollout, with access gradually expanding to paid subscribers and API users in the days following its 3 September launch.