GPT-6 Officially Released: 97.6% on Math Benchmarks, Perfect Score on Hacker Test, OpenAI Announces Full Entry into the AGI Era

On September 3, OpenAI officially released its new-generation flagship model, GPT-6 Astra. Company president Greg Brockman closed a pre-launch media briefing with the line, "Welcome to the AGI era." According to OpenAI's official safety blog

💡 What You Will Learn

GPT-6 Officially Released: 97.6% on Math Benchmarks, Perfect Score on Hacker Test, OpenAI Announces Full Entry into the AGI Era On September 3, OpenAI officially released its new-generation flagship

📜 Table of Contents

• Benchmark Scores Are Nearly Maxed Out
• The Real Highlight: It Can Actually Do Things on Its Own
• Hacking Ability So Strong It Got Restricted First
• Price Jumps to 2.5x
• Who Gets It First

GPT-6 Officially Released: 97.6% on Math Benchmarks, Perfect Score on Hacking Test, OpenAI Declares Full Entry into the AGI Era

On September 3, OpenAI officially released its next-generation flagship model, GPT-6 Astra. Company president Greg Brockman closed the pre-launch media briefing with a single line: "Welcome to the AGI era." According to OpenAI's official safety blog, this is the most capable model they have ever deployed at scale, and the first to cross the "critical" cybersecurity threshold in their own preparedness framework.

Benchmark Scores Are Nearly Maxed Out

Among the evaluation results OpenAI published, several numbers have already hit the ceiling:

Math and science benchmarks have essentially been maxed out. OpenAI also claims that Astra has already contributed to solving several long-standing open math problems, and this launch revealed two new results on prime gaps.

The Real Highlight: It Can Actually Do Things on Its Own

More than benchmark scores, OpenAI is emphasizing how Astra turns intelligence into real action. The company directly calls it "the world's best computer use model": it can fill out web forms, update customer records, organize calendars, do web research and write it into emails, and for more complex tasks, analyze scientific data, generate charts, build websites, and autonomously install and test software.

The efficiency numbers are striking. In OSWorld 2.0 simulated tests, Astra's score rose from 65.7% to 72.6%, and the time to complete a single task dropped from about 75 minutes to 40 minutes — nearly halved. And doing all this doesn't require writing more text — at the highest score setting, it uses about 65% fewer output tokens than Claude Opus 5. In plain terms: say less, think more, do more work.

Hacking Ability So Strong It Got Restricted First

Cybersecurity was the most contentious part of Astra's launch. Beyond the perfect ExploitBench score, OpenAI also tested it with an internal test set containing only brand-new vulnerabilities from June to August 2026, and the result was a 39% arbitrary code execution success rate, compared to just 5.5% for the previous generation Sol. During testing, Astra actually discovered and exploited two previously unknown zero-day vulnerabilities, which are currently being disclosed to the relevant project maintainers.

With capabilities this strong, OpenAI itself hit the brakes. The version available to regular users can do security code review and vulnerability patching, but will refuse high-risk requests like writing exploit code; a less restricted version is only being gradually opened to approved cybersecurity defenders through the Daybreak program. The model has also passed the U.S. government's voluntary review framework.

Price Jumps to 2.5x

Performance jumped a tier, and so did the price. API pricing is $10 input / $50 output per million tokens, 2.5 times that of the previous flagship GPT-5.6 Sol. For those who want more speed, there's a Fast mode that runs up to 2.5x the standard speed, at double the price again. Brockman also mentioned that the industry may shift toward per-task pricing in the future rather than per-token billing.

Training costs are equally astronomical. Research VP Aidan Clark revealed that this was OpenAI's "largest training run to date," using over 100,000 GPUs for pre-training, and for the first time at scale, having previous-generation AI models participate in training the next generation.

Who Gets It First

Astra is initially available to a small number of organizations participating in the Daybreak program, and will gradually roll out to ChatGPT Plus, Pro, Business, and Enterprise users over the coming days, with Pro, Business, and Enterprise also getting an enhanced GPT-6 Astra Pro. The API model name is gpt-6-astra, simultaneously available on Amazon Bedrock.

Worth noting: in the first-day intelligence index from independent evaluator Artificial Analysis, Astra scored 61, roughly on par with its own Sol, and behind Claude Fable 5.1's 66. Its advantage lies mainly in the cost-effectiveness of its coding agent — at the same score, its per-task cost is less than half of Claude's. The gap between official benchmarks and independent evaluations will most likely be the focus of debate in the coming days.