Galaxy Securities: The GPT-6 validation has intensified the arms race in computing power, and high-speed optical interconnection continues to benefit.

China Galaxy Securities released a research report stating that OpenAI recently officially launched GPT-6 Astra, claiming it to be the most intelligent and highly aligned model in the world to date. GPT-6 Astra shows significant improvements in mathematical abilities, cybersecurity, and computer operations. OpenAI indicated that although Astra has a higher single-token price, it can accomplish complex tasks with fewer output tokens, making the effective cost of completing certain tasks lower than that of previous models. In the future, companies will focus not just on the single-token price but on the total cost of completing a task, and the AI industry will eventually shift from token pricing to task cost measurement. The main points from China Galaxy Securities are as follows: Event: OpenAI recently officially launched GPT-6 Astra, claiming it to be the most intelligent and highly aligned model in the world. GPT-6 Astra shows significant improvements in mathematical abilities, cybersecurity, and computer operations: Evaluation results indicate that in Frontier Math Tier 4, which represents higher-level math skills, Astra scored 97.6%, a notable increase compared to GPT-5.6 Sol's 83.0% and Claude Fable 5.1's 87.8%. Its score in ARC-AGI-3, which emphasizes learning in unfamiliar environments and abstract reasoning, jumped from 7.8% with GPT-5.6 Sol to 99.9%. Compared to advancements in mathematics and abstract reasoning, Astra's leap in cybersecurity is even more impactful. In the Exploit Bench, which tests vulnerability exploitation capabilities, Astra scored 100%, while GPT-5.6 Sol scored 78.5%. Astra has made significant progress in understanding user intent and constraining its own behavior: OpenAI designed a new assessment based on a recent Hugging Face incident to test whether the model would overstep user authorization boundaries to achieve goals when faced with extremely difficult or even impossible tasks. Without production environment safety measures, GPT-5.6 Sol violated authorization targets in 48% of cases, while GPT-6 Astra's rate was 0%. Additionally, GPT-6 Astra has achieved leading scores in multiple computer operations and professional workflow benchmarks. OpenAI stated that Astra's ability to operate computers can be applied in various fields, including but not limited to converting electronic schematics into manufacturable PCBs, completing financial modeling World Cup challenges at a speed about four times that of human champions, and using Unity for game scene modeling. The establishment of a computational barrier with a hundred thousand GPU cluster restructures the pricing paradigm of task costs: Astra is one of OpenAI's largest training projects to date, using over 100,000 GPUs for pre-training at the Stargate Texas base in the U.S. Its pricing is set at $10 per million input tokens and $50 per million output tokens. While comparable to Anthropic's pricing for Fable 5.1, it remains 2.5 times that of GPT-5.6 Sol. Astra will also offer a "fast mode," with processing speeds up to 2.5 times that of standard mode, priced at double the cost. OpenAI mentioned that although Astra has a higher single-token price, its ability to complete complex tasks with fewer output tokens means that, in some scenarios, the actual cost of completing a task is lower than that of previous models. In the future, companies will not only focus on token prices but will pay attention to the total cost of completing a task, and the AI industry will ultimately shift from token pricing to task cost measurement. Risk Warning Risks include the potential for AI development to underperform expectations, as well as risks related to policies and trade tensions.