OpenAI released GPT-6 Astra on September 3, 2026, launching deployment the following day to a limited number of organizations across paid subscription tiers and enterprise channels. According to verified testing data, the new model achieves a 72.6 percent success rate on multi-step computer navigation tasks within an average of 40 minutes, marking a substantial performance leap over previous generations.
OpenAI GPT-6 Astra Release and Availability
OpenAI introduced GPT-6 Astra on September 3, 2026, initiating a staged rollout on September 4. According to corporate deployment notices, the software is accessible via Plus, Pro, Business, and Enterprise plans, as well as through the OpenAI API and Amazon Web Services. Free tier accounts do not receive access to the new model at launch, though OpenAI states that previous generations benefit from a roughly 60 percent speed increase driven by Astra’s underlying optimizations.
During an interview with CNBC, Sam Altman, le patron d’OpenAI, described the software as introducing a new level of capability that altered his own daily workflow. However, independent analysts emphasize evaluating the release through verifiable benchmark metrics rather than executive commentary.
OSWorld Benchmarks and Multi-Step Software Control
The primary performance gains of GPT-6 Astra center on long-horizon tasks requiring sequential execution across multiple software applications. On the OSWorld 2.0 offline benchmark, which tests automated desktop environment navigation, Astra records a 72.6 percent success rate completed in approximately 40 minutes per task. This compares against the predecessor model’s 65.7 percent success rate, which required an average of 75 minutes, according to test results reported by The New Stack.
Pro Tip: When evaluating multi-step tasks, assign specific software environments and check intermediate outputs rather than expecting single-prompt completion for complex workflows.
Additional technical evaluations show significant progress in specialized domains. According to benchmark summaries, Astra reaches 64.6 percent on Terminal-Bench Science, up from 22.4 percent for the prior generation. On ExploitBench, the model scores 100 percent, while achieving 99.2 percent on SRE-Bench across four evaluation attempts.
Critical Cybersecurity Classification and Deployment Safeguards
OpenAI has designated GPT-6 Astra with a “Critical” cybersecurity risk classification under its internal evaluation framework. This marks the first time an OpenAI model has reached this threshold, which indicates capabilities to identify and independently develop exploits for hardened real-world systems, or to design end-to-end attacks from a primary objective.
This classification prompted a staged, gradual rollout rather than immediate public availability.
Frequently Asked Questions
Is GPT-6 Astra available on free ChatGPT accounts?
No. OpenAI has not announced free tier access for GPT-6 Astra. The model is restricted to Plus, Pro, Business, and Enterprise subscription tiers, alongside API and Amazon Web Services access.
What is the core improvement of GPT-6 Astra over older versions?
The primary upgrade involves multi-step execution and software navigation. According to OSWorld 2.0 evaluations, Astra completes tasks in 40 minutes with a 72.6 percent success rate, compared to 75 minutes and 65.7 percent for the previous model.
Why did OpenAI classify the model as Critical for cybersecurity?
Under OpenAI’s risk evaluation framework, the Critical designation applies because the model’s capabilities reach the threshold of identifying and developing executable software vulnerabilities in real-world systems.
Explore More: Read our comprehensive coverage on enterprise artificial intelligence integration and browse our ongoing analysis of language model benchmarks.