According to BlockBeats, Farside monitoring showed that the US spot Bitcoin ETF saw a net inflow of $174.6 million yesterday, marking the third consecutive trading day of net inflows.

According to Beating AI News, since OpenAI released GPT-6 Astra on September 3, benchmark data for multiple models has been continuously adjusted. Some modifications have made Astra perform better, while the performance of some competing models has declined, raising questions about AI "ranking manipulation" and the transparency of evaluations.

Among these, Astra's hallucination rate dropped from 4.2% to 2%, and GPT-5.6 Sol's dropped from 12.2% to 9.4%, before recovering to 4.2% and 12.2% respectively. In the math assessment, Anthropic's Fable 5.1 score also dropped from 87.8% to 78%, but has since rebounded to 83%; GPT-5.6 Sol's score dropped from 83% to 80.5%, before recovering to 83%.

Furthermore, Astra's score in the ARC-AGI-3 benchmark improved from 98.6% in the pre-release draft to 99.99% in the final version, and its programming benchmark score also slightly increased from 57.7% to 57.9%. OpenAI stated that benchmark results are affected by factors such as model version, tool configuration, inference level, and test runs, and this adjustment was made to ensure that the data more accurately reflects the model's best performance.

However, Stanford University researchers believe that frequently rerunning evaluations may involve so-called "benchmaxxing," which involves maximizing benchmark scores by adjusting test conditions. Industry insiders point out that as competition among AI models intensifies, evaluation data has become a crucial tool for measuring model capabilities and competing for market share. Improving the transparency and reproducibility of benchmark testing is receiving increasing attention.