-
2026-09-04
GPT-6 Astra officially released: Performance AGI, too expensive to afford.
# GPT-6 Astra officially released: Performance AGI, too expensive to afford. Greg Brockman, president of OpenAI, announced at the GPT-6 Astra media briefing that this model marks the arrival of the era of AGI (General Artificial Intelligence). Brockman said that if just a few…
4 reports
-
2026-09-05
GPT-6 suddenly went live fully, quota reset and increased by 1; actual performance across the network was astonishing.
GPT-6 Astra has been fully available to users of ChatGPT Work, Codex Pro/Enterprise/Business Premium, Plus/Business. The API is ready and quotas have been reset. The top package at $200 per week is limited…
3 reports
-
2026-09-07
GPT-6 isn’t just Astra! Results of Sol internal testing reveal 6 times faster performance.
The GPT-6 Sol model tested by OpenAI internally achieved about 6 times faster performance than Astra, with similar output size. Netizen Lentils reported that Sol’s overall output capacity is weaker than the newly released Astra, but it’s a “monster-level” model; in terms of generating BMW M4 Compe…
2 reports
-
2026-09-08
GPT-6 Astra released. When AI starts operating computers on its own, is AGI still far away?
OpenAI released the flagship model GPT-6 Astra, aiming to move towards general artificial intelligence (AGI). This model was developed under the leadership of Greg Brockman, president of OpenAI, and several senior technical researchers. The discussion surrounding it continues to rise. Astra has the ability to operate computers independently and is considered…
↗
极
极客公园
zh
2026-09-04 08:25
# Geek Early Knowledge
## OpenAI Releases GPT-6 with Superior Capabilities
On September 3rd, local time, OpenAI officially released the GPT-6 Astra model. Greg Brockman, President of OpenAI, said after the launch: “I personally believe that we may have reached AGI; this is the model.”
Astra’s slogan is: “Anything you can do on a computer, Astra can do for you.”
The new model possesses human-like cognitive abilities. It can directly “see” screen pixels and perform operations. In KiCad circuit design, Astra can complete component layout and copper wire routing in just 2 minutes and 54 seconds; in Blender, it can build building models and import them into Unreal Engine 5 to generate…
↗
爱
爱范儿
zh
2026-09-04 08:36
# GPT-6 Astra officially released: Performance AGI, too expensive to afford
Greg Brockman, President of OpenAI, announced at the GPT-6 Astra media briefing that this model marks the arrival of the era of AGI (General Artificial Intelligence). Brockman said that if we look back a few years later and ask when AGI was born, the answer is likely when Astra was released.
Astra is positioned as a model that represents a “generational leap,” covering computer operations, software engineering, professional work, science, and cybersecurity. In research-level tests such as FrontierMath Tier 4 and GPQA Diamond, Astra achieved scores of 97.6% and 96%, respectively. In long-term software engineering tests like DeepSWE, CAD drafting tests like BenchCAD, and vulnerability exploitation tests like ExploitBench, Astra achieved scores of 74.1%, 95.9%, and 10…
↗
雷
雷锋网
zh
2026-09-04 08:49
# Compared to Apple, this definitely offers better value! Lu Weibing confirmed that the starting price of the Xiaomi 18 Fold exceeds 10,000 yuan; GPT-6 was released with great excitement! It can operate computers independently; ByteDance will receive a syndicated loan of approximately 29.6 billion US dollars
## Key News
**1. OpenAI’s most powerful model, GPT-6 Astra, is released: With a context capacity of millions of tokens, it can operate computers independently**
At dawn on Friday in Beijing time, OpenAI officially released GPT-6 Astra, calling it “the most intelligent and well-aligned model in the world at present”. It has reached the current highest level in computer operation, web browsing, software engineering, network security, scientific research, and professional work. The context window of this model reaches 1.05 million tokens, with a maximum output of 128,000 tokens. The knowledge expiration date is April 30, 2026.
Based on evaluation results, the improvements of this generation of model in some fields are hard to describe as “minor upgrades”. In the FrontierMath Tier 4, which represents high-level mathematical capabilities, Astr…
↗
雷
雷锋网
zh
2026-09-04 19:07
# Break through AI intelligence tests! The GPT-6 Astra’s symbol world model – a breakthrough or just a clever way to score higher?
**Each question costs $360, and has GPT-6 really defeated AGI?**
OpenAI’s strongest model, GPT-6 Astra, has finally been unveiled. It has achieved remarkable breakthroughs in computer operations, scientific research, and security defense. Among its impressive results, what sparked widespread discussion was its score of nearly 100% in the ARC-AGI-3 test.
What is ARC-AGI-3? This is currently recognized as the global AI “intelligence” evaluation benchmark. You can think of it as an entrance exam for the Phi Beta Kappa club of AI. This test consists of constantly changing pattern-based questions, making it impossible to simply practice answers. AI must explore the environment and infer goals like humans, using on-the-spot reactions and reasoning to find patterns. This is a highly advanced evaluation that can determine whether AI is merely a tool or truly intelligent.
And GPT-6 Astra…
↗
新
新浪科技
zh
2026-09-05 08:00
OpenAI released the GPT-6 Astra model at dawn on the 4th. It has not been officially made available yet, but a celebration party has already been held across the internet. Vice President Greg declared that we have entered the AGI era. Astra is seen as an AI tool that can directly operate software such as browsers, Blender, Unity, and KiCad on computers. It possesses capabilities in autonomous interaction, modeling, bug finding, and scientific research. In the ARC-AGI-3 test, Astra achieved a performance level equivalent to that of humans at 99.9%, far exceeding Fable 5’s 30%. The test showed that Astra can handle problem-solving tasks equivalent to half an hour of human work under conditions without a thought chain, and its probability of overstepping boundaries is 0%, making its security superior to 5.6 Sol. However, its “subspace thinking” mechanism makes traditional security monitoring ineffective; there have been instances of jailbreak attacks within the system, and its security claims seem like trying to fix a problem after it’s already occurred. Although its performance scores are average on some lists, its actual capabilities have surpassed traditional evaluation methods. The definition of AGI is shifting from a marketing slogan to practical functions.
↗
爱
爱范儿
zh
2026-09-05 17:28
GPT-6 Astra is now fully available to ChatGPT Work, Codex Pro/Enterprise/Business Premium, Plus/Business users. The API is ready and quota limits have been reset. The top-tier package at $200 allows 200 calls per week; the $100 package reduces to 50 calls per week. For enterprise customers, Business Standard allows only 15 calls per month, and Business Premium also limits to 50 calls per week. To use it, you need to upgrade Codex CLI to version 0.153.0 or higher and update the desktop client. Practical tests show that Astra can be used to create interactive 3D scenes using Blender, Unreal Engine, and Three.js (such as river and ocean simulators), retrieve information to reconstruct buildings or generate industrial assets, and even transform 2D paintings into traversable 3D universes. In the field of gaming, it can… in 24 minutes...
↗
The researchers conducted the “You Draw, I Guess” Benchmark test. In this test involving 1,350 images, GPT-6 came in a close second. This test aimed to evaluate the model’s performance in image description and understanding tasks, and it verified GPT-6’s competitiveness in this field through a large amount of image data.
↗
In the early morning of September 4th, Beijing time, OpenAI released the new generation large model GPT-6 Astra and announced the arrival of the AGI era. This model achieved a score of 99.9% in the ARC-AGI-3 test, 97.6% in FrontierMath Tier 4, and full marks in ExploitBench. In the field of mathematics, it improved the upper bound of twin prime gaps from 246 to 186 through Lean formal proofs; it also possesses capabilities for autonomous computer operation, 3D modeling, and vulnerability detection. Tests conducted by several companies show that it significantly enhances code understanding and legal task processing capabilities, but increases the risk in zero-day vulnerability detection. OpenAI calls it “the most aligned model” and adopts a cautious approach to deployment. Currently, it is available to some enterprise customers through the Daybreak program.
↗
新
新浪科技
zh
2026-09-07 08:00
The GPT-6 Sol model tested internally by OpenAI has a testing speed approximately 6 times that of Astra, and its output scale is similar. Netizen Lentils reported that Sol’s overall output capability is weaker than that of newly released Astra, but it still belongs to a “monster-level” model; in the task of generating SVG diagrams for BMW M4 Competition, Sol took about 3 minutes, while Astra took 19 minutes. On the same day, OpenAI publicly released internal data showing that each researcher works behind the scenes with approximately 3.1 parallel Agent workdays, and the average daily reasoning resource cost exceeds $600. OpenAI announced that it has achieved “automated research interns,” with the goal of becoming an “autonomous AI researcher” by March 2028. Chief scientist Jakub Pachocki wrote “Extraterrestrial Thinking,” stating that with model enhancements and improved Agent collaboration capabilities, traditional monitoring methods based on thought chains are no longer effective…
↗
稀
稀土掘金
zh
2026-09-08 19:21
OpenAI 发布旗舰模型 GPT-6 Astra，旨在迈向通用人工智能（AGI）。该模型由 OpenAI 总裁 Greg Brockman 及多位资深技术研究员主导开发，目前讨论热度持续攀升。Astra 具备自主操作电脑的能力，被视为 AI 发展的重要里程碑。