“GPT-5对人类的阿谀奉承减少了”

Core Insights - OpenAI has launched GPT-5, claiming it to be the most intelligent and fastest model to date, with advanced capabilities in various fields such as programming, mathematics, writing, health, and visual intelligence [2][3] - GPT-5 is available to all users, with Plus subscribers gaining access to the pro version sooner [2] - The model has shown significant improvements in coding, writing, and reasoning capabilities compared to its predecessors [3][5] Performance Metrics - In benchmark tests, GPT-5 scored 94.6% and 100% in AIME2025 tests, outperforming previous models [4] - In expert-level math tests, GPT-5 scored 13.5% and 32.1%, while in PhD-level science questions, it scored 85.7% and 89.4% [4][5] - GPT-5 demonstrated superior performance in software engineering and multi-language code editing tests, scoring 74.9% and 88% respectively [5] Improvements and Features - GPT-5 has reduced the number of tokens generated by 50% to 80% in various reasoning scenarios, indicating more efficient output [6] - The model has a lower hallucination rate compared to previous versions, with a 45% reduction in factual errors during web searches [6] - OpenAI has reduced the tendency of GPT-5 to flatter users, with the probability of such behavior dropping from 14.5% to below 6% [6] Pricing and Market Position - GPT-5 offers competitive pricing for API services, with input and output costs lower than previous models [7] - The time gap between the launches of GPT-4 and GPT-5 is approximately two and a half years, indicating a slower update pace [7][8] - Despite the advancements, some benchmark scores of GPT-5 are not significantly different from earlier models, raising questions about its position in the AI landscape [8] Industry Reactions - The launch of GPT-5 has drawn mixed reactions, with some industry figures expressing pride in competing models outperforming GPT-5 in specific benchmarks [8]