Galaxy Securities: GPT-6 verifies computing power, the arms race heats up, high-speed optical connectivity continues to benefit

The Zhitong Finance App learned that China Galaxy Securities released a research report saying that recently, OpenAI officially released the GPT-6 Astra, and called it currently the smartest and most aligned model in the world. GPT-6 Astra has shown significant improvements in math skills, network security, and computer operation performance. OpenAI said that although Astra's single token is more expensive, since it can complete complex tasks with fewer tokens, the actual cost of completing a task is lower than the previous model in some scenarios. In the future, enterprises will not only focus on the unit price of tokens, but will focus on the total cost of completing a task. Eventually, the AI industry will shift from token pricing to task cost measurement.

The main views of China Galaxy Securities are as follows:

Incident: Recently, OpenAI officially released GPT-6 Astra and called it currently the smartest and most aligned model in the world.

GPT-6 Astra showed significant improvements in mathematical ability, network security, and computer operation performance: Judging from the evaluation results, Astra's score reached 97.6% on Frontier Math Tier 4 representing advanced mathematical abilities, which is a significant improvement compared to 83.0% of GPT-5.6 Sol and 87.8% of Claude Fable 5.1; in ARC-AGI-3, which emphasizes learning and abstract reasoning in unfamiliar environments, its results are even more from GPT- 5.6 Sol's 7.8% jumped to 99.9%. Compared to math and abstract reasoning, Astra's rise in cybersecurity is more realistic and impactful. In the Exploit Bench, which tests exploitability, Astra scored 100%, while GPT-5.6 Sol scored 78.5%. Astra has made significant progress in understanding user intent and restraining its own behavior: Referring to the recent Hugging Face incident, OpenAI has designed a new evaluation to test whether the model will break through user authorization boundaries in order to achieve goals when faced with extremely difficult or even impossible tasks. In the absence of production environment safety measures, GPT-5.6 Sol will break the authorized target in 48% of cases, compared to 0% for GPT-6 Astra.

Additionally, the GPT-6 Astra achieved leading results in several computer operation and professional workflow benchmarks. OpenAI said that Astra's ability to operate computers can be used for output in various fields, including but not limited to turning electronic schematics into manufacturable PCBs, completing financial modeling World Cup challenges at about four times the speed of human champions, and modeling game scenarios using Unity.

The 100,000 card cluster establishes computing power barriers and restructures the pricing paradigm: Astra is one of OpenAI's largest training programs to date. The model uses more than 100,000 GPUs for pre-training at the US StarGate Texas base. It is priced at $10 per million input terms and $50 per million output terms. Although the price of Fable 5.1 is the same as Anthropic, it is still 2.5 times that of GPT-5.6 Sol. Astra will also offer a “fast mode” that can process up to 2.5 times faster than the standard mode and double the price. OpenAI said that although Astra's single token is more expensive, since it can complete complex tasks with fewer tokens, the actual cost of completing a task is lower than the previous model in some scenarios. In the future, enterprises will not only focus on the unit price of tokens, but will focus on the total cost of completing a task. Eventually, the AI industry will shift from token pricing to task cost measurement.

Risk Alerts

The risk that AI development falls short of expectations, the risk of policy and trade friction, etc.