GPT-6 Astra is OpenAI’s new frontier model, and its arrival brings two questions to the center of the AI market: what these systems can do, and what they cost to run. OpenAI describes Astra as its most capable and most aligned model so far. Access opened first to a limited set of organizations, then widened to paid users across ChatGPT Plus, Pro, Business, and Enterprise. The GPT-6 Astra release date fell in early September, with a limited preview ahead of broader access.
Developers can reach the model through the OpenAI API, Microsoft Azure, and Amazon Bedrock. The company built Astra for long, multi-step work rather than short chat. That design shows in the strongest gains, which sit in GPT-6 Astra computer use. OpenAI says the model can fill out online forms, update customer records, organize a calendar, run research, and draft summaries inside a user’s email or document editor. The aim is a system that finishes tasks, not one that only answers questions.
- ADX has activated its SWIFT Business Identifier Code, ADSEAEADXXX, and joined the global SWIFT network.
- The link lets ADX trade standardised financial messages with custodians and securities depositories worldwide.
- The step strengthens ADX’s post-trade infrastructure and its pull for international investors.
- ADX ties the move to Abu Dhabi’s push to be a leading international financial centre.
- ECONOMY
What the GPT-6 Astra benchmarks show
The GPT-6 Astra benchmarks published by the company are aggressive. OpenAI says the model saturates FrontierMath Tier 4 at close to 98 percent, reaches 99.9 percent on ARC-AGI-3, and scores 100 percent on ExploitBench. Both the math and reasoning tests were designed to stay ahead of AI systems, so results this high point to a real step up. One caveat matters here. The ARC-AGI-3 score used a general-purpose setup that preserved the model’s reasoning and managed long context, so it is not a clean match with every earlier figure.
Independent testing gives a more measured read. Artificial Analysis found Astra roughly level with its predecessor on overall intelligence, and behind Claude Fable 5.1 on general reasoning, while gaining on coding at lower cost.
On computer-use speed, OpenAI reports a higher score on the OSWorld 2.0 test in about 47 percent less time per task than the prior model, GPT-5.6 Sol. For firms weighing automation of desk work, time per task can weigh as much as raw accuracy.
Cost and cybersecurity set the real test
GPT-6 Astra pricing is where the tradeoff shows most clearly. Standard rates run at 10 dollars per million input tokens and 50 dollars per million output tokens, about 2.5 times the prior flagship’s 4 and 20 dollar rates. OpenAI notes that Astra often uses fewer tokens for similar work, which can offset part of the higher rate. Whether that holds depends on the task. The number worth tracking is cost per finished result, not the headline token price.
On safety, Astra is the first OpenAI model to reach the Critical level for cybersecurity under the company’s internal framework. OpenAI says the model can find unknown security flaws and build ways to exploit them across well-defended systems without a person guiding each step. To limit misuse, the company restricts the most advanced exploit abilities and adds production safeguards. It also flagged a weakness in how well its monitors can read the model’s reasoning under pressure, and named that as open work.
The wider shift is about where value moves. As frontier models take on full tasks rather than single answers, the market question moves from capability alone to capability set against cost and risk. The OpenAI GPT-6 Astra release, arriving alongside strong models from rival labs, brings that calculation into the open for enterprise buyers.
