Controversial! The US's Strongest Model Opus 4.8 Is Accused of Being a Distilled Version of a Chinese Open-Source Large Model

Anthropic, a leading company in the U.S. large model industry, officially released the Claude Opus 4.8 large model at midnight today. However, while the model achieved outstanding results on multiple rankings, it has been reported by many netizens and developers to have "cognitive confusion" issues.

A netizen found through an API test that when asking the model about its identity, it mistakenly claimed to be Alibaba's "Qwen" or "DeepSeek," two major open-source large models in China.

API Testing Triggers Cognitive Abnormality

Regarding this phenomenon, some users testing on the web interface said they could not reproduce it and believed the model had not distilled Chinese large models. However, relevant developers explained that the web interface conversation had strict system prompts, which concealed this anomaly.

In contrast, in low-level API testing without any constraints, this issue was frequently reproduced. This directly indicates that the U.S. top-tier large model likely used a large amount of generation data from Chinese open-source large models for "distillation" during training or fine-tuning.

Inconsistent Behavior Raises Double Standards Criticism

This incident caused a big stir in the technology community, mainly because Anthropic previously showed a strong "anti-China" tendency. The company once collaborated with the U.S. Department of Defense and other departments to draft legislation, attempting to define "large model distillation" as an adversarial attack.

They not only called on the U.S. government to increase chip and software restrictions on China's AI industry but also proposed a strategic goal to achieve a two-year lead over China in AI by 2028. Now, their flagship model was caught distilling Chinese large models, ironically exposing their double standards.

The Largest Chip Leasing Transaction in History Has Emerged! Apollo Joins Blackstone to Raise $36 Billion to Aggressively Acquire Google TPU for Anthropic

A capital frenzy is sweeping global artificial intelligence infrastructure development. Apollo Global Management and Blackstone Group plan to jointly launch a debt financing of approximately $36 billion, providing financial support to AI unicorn Anthropic through a 'rent-to-own' model. This will become one of the largest private credit transactions in history, accelerating the deployment of top-tier AI infrastructure.

Earn 47 Billion Dollars! Anthropic's Annual Revenue Surges, Claude Enterprise Applications See a Major Breakthrough

Anthropic announced that its annual revenue has exceeded 47 billion dollars, demonstrating the huge potential of generative AI in commercialization. Since the Series B funding round in February, the flagship model Claude has seen a major breakthrough in the enterprise market, driving rapid growth in performance and highlighting the intensifying competition in the AI industry.

Rushing Before Listing! Anthropic Raises $6.5 Billion with a Valuation Close to $1 Trillion

AI company Anthropic completed a $65 billion Series H funding round, with a post-investment valuation of $965 billion. Led by Altimeter Capital and Sequoia Capital, with participation from Samsung and others, this is seen as the final private round before IPO. The funds include prior commitments from tech giants like Amazon.....

Anthropic Completes $6.5 Billion Series H Funding, Launches Claude Opus 4.8 Model, and Valuation Approaches $1 Trillion

Anthropic completes $6.5 billion Series H funding, with a post-money valuation of $965 billion, approaching $1 trillion. The round was led by Altimeter Capital, with participation from Samsung, SK Hynix, Micron, and Amazon. The funds will be used for AI safety and explainability research and scaling up.

Code Flaws Reduced by 75%! Anthropic Launches Claude Opus 4.8, Speed Increased by 2.5 Times, Outperforming Industry Competitors

Anthropic releases a refined version of its flagship model, Claude Opus 4.8, focusing on agent programming, multi-domain reasoning, and knowledge work. This version surpasses GPT-5.5 in multiple core benchmark tests, significantly reducing the phenomenon of 'AI lying,' with a sharp decrease in programming defects and improved judgment, marking a major breakthrough in AI code reliability.