OpenAI has officially launched GPT-6 Astra, positioning it as the company’s most capable model for professional workflows. Unlike previous iterations that demanded heavy data preparation and custom API integrations, Astra is designed to interact directly with existing desktop applications. This “computer use” capability allows the model to navigate software without dedicated APIs, letting businesses deploy AI into their current stacks from day one. Early adopters are already leveraging this flexibility for tasks ranging from financial discrepancy detection to GPU optimization.
The internal testing results at OpenAI were striking. Engineering teams used Astra to identify and resolve a persistent memory-allocation bottleneck in Codex test environments. By switching allocators, they achieved a 25x reduction in turn latency while handling roughly 30% higher peak memory usage. Marketing teams also utilized the model to edit three hours of multicamera footage into a launch video, which amassed over 550,000 views in just four days. These examples highlight a shift from simple text generation to complex, multi-step task execution across different media and codebases.
Cost efficiency remains a central pillar of the release. Astra is trained to complete tasks using fewer tokens and requiring fewer retries, directly lowering the cost per task for enterprise clients. Pricing starts at $10 per million input tokens and $50 per million output tokens. OpenAI claims this places Astra at the forefront of the cost-efficiency frontier for professional work and coding, outperforming competitors on benchmarks like Terminal Bench 4.0. For businesses watching their burn rate, this reduction in rework could make high-frequency AI usage financially viable where it previously wasn’t.
Safety features have been tightened to address the risks of autonomous agent behavior. Astra produced unintended outcomes 89% less often than GPT-5.6 Sol in internal safety benchmarks involving sensitive data handling. It is also the first model to reach the Critical cybersecurity capability threshold under OpenAI’s Preparedness Framework. With new enterprise admin controls allowing restrictions on specific websites and applications, organizations can gradually expand access. The real test will be whether these guardrails hold up when Astra interacts with legacy systems that lack modern security protocols.
