- Published on
Has the AGI era truly arrived?
- Authors
- Name
- DeepHorus
OpenAI officially released its latest flagship model, GPT-6 Astra, in the early hours of September 4th, Beijing time. The company's president, Greg Brockman, declared at a pre-release media briefing, "Welcome to the era of AGI (Artificial General Intelligence)." This release followed closely after a widespread outage affecting major AI services like ChatGPT, Claude, and Grok on the evening of September 3rd. While netizens joked that the outage was related to Astra, there was no evidence to support this claim.
According to the release notes, Astra is not yet fully rolled out. It is currently available only to a limited number of institutions within the OpenAI Trusted Access / Daybreak system and to vetted cybersecurity users. Plus, Pro, Business, Enterprise subscribers, and API users are expected to gain access over the coming days. Aidan Clark, OpenAI's Vice President of Research, noted that this is the first time the company has used over 100,000 GPUs for pre-training at its Stargate data center in Texas, and it is the first model to deeply involve previous-generation models in the training process.
In benchmark testing, Astra performed exceptionally well, particularly on the ARC-AGI-3 benchmark designed to evaluate a model's ability to explore unknown environments. Under OpenAI's own Provider Adapter configuration, Astra achieved a top score of 99.9%. When using the standardized Standard harness, its score was 62.7%. OpenAI explained that by optimizing configurationsāincluding preserving reasoning and compression settingsāthe model's efficiency in production systems is greatly enhanced, running approximately 3.66 times faster while consuming 49% fewer tokens. In other tests, Astra scored 97.6% on FrontierMath Tier 4 and achieved an 88.0% success rate on the first attempt in the SRE-Bench reverse engineering benchmark.
The release of Astra marks a shift in OpenAI's product focus, moving from what a model can "answer" to what it can directly "do." The official website showcases its capabilities in complex tasks across multiple domains, including KiCad, Excel, Blender, browser operation, game development, and data analysis. OpenAI even calls it "the world's best computer use model." Technically, Astra introduces a new context mechanism that retains notes and allows for information backtracking when tasks are paused, enhancing its ability to handle long-range, multi-step tasks.
However, these enhanced capabilities come with a price increase. The standard API pricing for GPT-6 Astra is: $10 per million input tokens, $1 for cached input, and $50 for output. This represents an increase of approximately 2.5 times compared to the previous flagship model, GPT-5.6 Sol. OpenAI positions Astra as the flagship model for complex tasks, while cost-sensitive users can continue using the GPT-5.6 Terra or Luna models.
In terms of safety alignment, Astra shows significant improvements. OpenAI's internal testing indicates that its violation rate dropped from 22.0% for Sol to 2.4%, and its hallucination rate decreased from 9.4% to 2.0%. However, OpenAI's Chief Scientist, Jakub Pachocki, also warned that "progress in intelligence does not guarantee progress in alignment." Because Astra has enhanced control over its own chain of thought, reducing explicit reasoning steps, this poses challenges for traditional monitoring methods. To address this, OpenAI has deployed a new "misalignment monitoring" system in production, which comprehensively observes the model's reasoning, actions, and task trajectories.
Looking ahead, the release of Astra not only demonstrates major advances in AI's ability to execute complex tasks and maintain long-term memory but also reveals new paradigms required for safety monitoring under more powerful models. Although full access is still some time away, this modelāpositioned as a crucial step toward the AGI eraāhas already sparked in-depth discussions within the industry about the boundaries of AI capabilities and the governance of safety.
