实测GPT-5：写作坠入谷底，编程一骑绝尘。

Core Viewpoint - The article discusses the launch of GPT-5, highlighting its advancements over previous models and the implications for AI development and user interaction [2][9][16]. Model Overview - GPT-5 is a unified system that includes a fast model for general queries (gpt-5-main) and a deep reasoning model for complex questions (gpt-5-thinking) [11]. - The system utilizes a real-time router to dynamically select the appropriate model based on conversation type, complexity, and user intent [12][14]. - Additional models include mini versions for handling excess requests and a Pro version for parallel computing [15][14]. Performance Improvements - GPT-5 significantly reduces factual inaccuracies, with gpt-5-main producing 44% fewer major factual errors compared to GPT-4o, and gpt-5-thinking achieving 78% fewer errors than OpenAI o3 [19][20]. - In benchmark tests, GPT-5 models show a substantial decrease in hallucination rates, with gpt-5-thinking producing five times fewer factual errors than OpenAI o3 [22]. - The model also exhibits improved handling of sycophancy, with a 69% reduction in such behavior among free users and 75% among paid users compared to GPT-4o [24][27]. Benchmarking and Rankings - GPT-5 achieved top scores across various assessments, including math competitions and multi-modal capabilities, outperforming previous models [30][43]. - It ranked first in the latest large model blind test rankings, demonstrating superior performance in multiple categories [45]. Energy Efficiency - GPT-5 is noted for being more energy-efficient, with a 50-80% reduction in output tokens used for tasks like visual reasoning and programming [47][48]. Developer Pricing - The pricing for developers using GPT-5 is set at $1.25 per million tokens (with a 90% caching discount) and $10 per million tokens for output [54]. User Experience - Initial user feedback indicates mixed results, with some users noting that GPT-5's writing and emotional intelligence may not surpass that of GPT-4.5 [59][68]. - However, GPT-5 has shown strong performance in production-level coding tasks, indicating its potential for practical applications [99].