I couldn’t answer questions such as washing a car while walking, red-green color blindness, etc. Netizens complained that the most powerful programming AI has seriously degraded my intelligence.

Although Mythos, released by Anthropic a few days ago, has set a new record for the most powerful AI large model, it is not open for use. The current most powerful AI programming large model is Claude Opus 4.6, which is also the most expensive currently. However, the recent problem with Claude’s intelligence has caused dissatisfaction among a large number of developers who paid high prices.Previously, even AMD's AI department executives came out and said that the problem of intelligence reduction has affected development, and the scope of the impact is quite large.

In the domestic Linux.do community, many netizens also talked about the problem of Opus 4.6's intelligence reduction, and also verified it with two well-known AI large model test questions.

One of these two questions is about whether to drive or walk to the car wash at a distance of 50 meters. The other is about the father's breakdown because of his daughter's red-green color blindness.In essence, they are all logical traps, but they can test the ability of AI.

Many domestic and foreign AIs have overturned on the car washing problem before. Claude's large model is one of the few large models that can answer the question correctly, with obvious IQ advantages.

However, according to current actual measurements, even the most powerful Opus 4.6 large model has begun to answer these two questions incorrectly, and has lost its previous high IQ ability.

Some netizens also posted pictures of their own tests, saying they could answer it correctly, so this question is not completely one-sided.

Judging from AMD's previous complaints, it is clear that Cluade's intelligence has recently been reduced, but the reason for the reduction is difficult to say. Anthropic's executives did not admit that there was a problem before.

A more reasonable guess is that it is related to computing power. Musk's question a few days ago inadvertently revealed that Claude's Sonnet 4.6 is a large model with 1 trillion parameters.Opus 4.6 has 5 trillion parameters.And it is not of the MoE type. The dense model requires very high computing power and is costly, so the charging price is also the most expensive.

Today's AI companies are still losing money selling subscriptions, which can also partially explain Anthropic's motivation to ban applications with extremely high token consumption like Crayfish.

For domestic developers, there is another reason to reduce intelligence.That is, Anthropic will target intelligence based on IP.Essentially, it is also a matter of allocating computing power, but they are not the only ones to do this in the form of targeted IP/country. Other American companies such as OpenAI have also been found to have used this method to restrict domestic developers and users.