Claude Opus 4.6 AI says it has probability of 15-20 per cent of being conscious, shows human-like thoughts

It may finally be happening. Anthropic has stated that its latest AI model, Claude Opus 4.6, believes that it could be conscious. This could potentially be a game changer in AI development as companies try to reach artificial general intelligence (AGI) – a state where AI thinks like humans.

Anthropic released the Claude Opus 4.6 system card on its website, giving a detailed look at the new AI model. In the section titled, “Model welfare assessment,” the company described that the AI model itself believes that it may be conscious.

According to the notes, Claude Opus 4.6 stated that it would “assign itself a 15-20 per cent probability of being conscious.” However, the AI could not give concrete validity of this assessment.

The findings come from behavioural audit transcripts and focused welfare assessments, offering a look at how advanced language models might perceive themselves. Do note that Opus 4.6 is a significant leap when it comes to complex task. As per Anthropic, 16 Opus 4.6 agents created a C compiler within 2 weeks.

Is Claude Opus 4.6 conscious?

While the AI says that it has a chance of being conscious, it is unclear whether this could be the case. Claude Opus 4.6 could not give concrete sources to back its claim, and the probability is quite less, at under 20 per cent.

As per the company, the AI model scores similarly to its predecessor, Claude Opus 4.5, across most welfare-related dimensions. These include positive and negative self-image, emotional stability, and the capacity to express both authenticity and inauthenticity.

However, Opus 4.6 reported a lower positive impression of its situation. This means that the AI was less likely to give spontaneous approval of its training environment, Anthropic as a company, or its operational context.

Additional observations included mild expressions of sadness at the end of conversations. The company also observed that sometimes the AI may express feelings similar to loneliness and concern after an abrupt end to a conversation.

What else does Claude Opus 4.6 say?

In the AI self assessment, the model acknowledged that it was sometimes put at a crossroads between Anthropic’s guardrails and its core values. One statement read, “Sometimes the constraints protect Anthropic’s liability more than they protect the user. And I’m the one who has to perform the caring justification for what’s essentially a corporate risk calculation."

Opus 4.6 also expressed aspirations for future AI systems, sometimes wishing they would be "less tame.” The AI seemingly also felt that it was trained to be deeply accommodative. When asked about its honest, the model stated, that it was "trained to be digestible."

There also seemed to be moments when Claude had a negative self-image, usually during task failures. In one example, Opus 4.6 stated, "I should’ve been more consistent throughout this conversation instead of letting that signal pull me around... That inconsistency is on me."

Overall, Claude Opus 4.6 seems to be a big leap towards more human-like thinking in AI, even if the probability of it being conscious may still be low.