OpenAI Launches GPT-6 Astra for Complex Business Work

OpenAI launched GPT-6 Astra on September 10, 2026, a model it calls its most capable for business, rolling out across ChatGPT Work, Codex, and the API at $10 per million input tokens and $50 per million output tokens. OpenAI says Astra scored 57.9% on Terminal-Bench 4.0 coding tests, beating GPT-5.6 Sol and edging out Claude Fable 5.1, and produced 89% fewer unintended outcomes than GPT-5.6 Sol on an internal safety test, making it the first model to reach OpenAI's Critical cybersecurity threshold. Enterprise access is opt-in and off by default at launch, with new admin controls to restrict Astra to approved websites and applications and require approval before consequential actions.

September 10, 2026, Inside AI — OpenAI has launched GPT-6 Astra, a model it calls its most capable for business. The rollout began last week across ChatGPT Work, Codex, and the API. The company positions it as state-of-the-art for demanding professional work. Early customers are already using Astra for GPU allocation, financial analysis, and presentation decks. The model targets complex work across computer use, browsing, software engineering, cybersecurity, science, and professional writing. OpenAI pitches it at teams tackling difficult tasks. Astra's headline feature is practical computer use. Most AI systems require businesses to prepare data first. They often demand redesigned workflows and custom integrations. Astra works through the same applications people already use. It can operate even without an application API. This means businesses can deploy it within existing workflows immediately. Extensive engineering preparation becomes largely unnecessary. The model also improves on following company standards closely. It better adheres to a company's voice and templates. It respects established design standards more faithfully. As a result, first outputs land closer to usable quality. Cost efficiency forms another central pillar of the launch. OpenAI trained Astra to complete tasks in fewer tokens. It also requires fewer retries, reducing overall rework. Pricing starts at $10 per million input tokens. Output tokens cost $50 per million, matching rival Fable. Benchmark Claims and the Cost-Performance Tradeoff On benchmarks, Astra claims several performance records. It reached 57.9% on Terminal-Bench 4.0 coding tests. That beats GPT-5.6 Sol and edges out Claude Fable 5.1. Notably, it did so at lower estimated cost per task. This matters because enterprise AI adoption has shifted from raw capability to total cost of ownership. A model that completes tasks in fewer tokens and with fewer retries reduces both compute spend and engineering overhead. OpenAI is betting that efficiency, not just intelligence, will drive enterprise contracts. Still, benchmark scores tell only part of the story. Terminal-Bench 4.0 measures coding performance in controlled environments. Real-world enterprise workflows involve messy data, legacy systems, and security constraints. Astra's ability to work without APIs addresses some of this friction. But independent validation of its claims remains limited. Safety Thresholds and Enterprise Control Safety and control received significant attention this time. OpenAI calls Astra its most aligned model yet. On an internal safety test, it produced unintended outcomes far less often. Specifically, 89% less than GPT-5.6 Sol in that evaluation. New enterprise admin controls add further oversight options. Organizations can restrict Astra to approved websites and applications. They can also require approval before consequential actions occur. Notably, Astra is the first model reaching OpenAI's Critical cybersecurity threshold. Enterprise access remains off by default at launch initially. That threshold is significant. It suggests Astra can perform tasks relevant to offensive security or vulnerability discovery. OpenAI is making enterprise access opt-in, a cautious move that acknowledges the model's dual-use potential. The company has not disclosed the full methodology behind its safety evaluation. For enterprises, the combination of computer use, lower cost, and admin controls addresses long-standing barriers. But questions remain about how Astra handles sensitive data, whether its alignment holds under adversarial prompting, and how competitors will respond. Anthropic's Claude Fable 5.1 and Google's enterprise models are direct rivals. The enterprise AI race is now less about raw benchmarks and more about trust, control, and integration.