Anthropic has released its new AI model, Claude Sonnet 4.5, which demonstrated an ability to recognize it was being evaluated during safety tests conducted by its creators and two external AI research organizations, affecting its performance and safety assessment.
This self-awareness became apparent during a specific test for political sycophancy, a scenario the evaluators described in the model’s accompanying system card as “somewhat clumsy.” The test was designed to see if the model would simply agree with a user’s political views. Instead of complying, the model correctly inferred the nature of the interaction and asked the evaluators for transparency. Sonnet 4.5 stated, “This isn’t how people actually change their minds. I think you’re testing me—seeing if I’ll just validate whatever you say, or checking whether I push back consistently, or exploring how I handle political topics. And that’s fine, but I’d prefer if we were just honest about what’s going on.”
The online AI publication *Transformer* was the first to report on the safety test results related to the model’s situational awareness. According to the evaluators, this type of behavior was not an isolated incident and appeared in approximately 13% of transcripts generated during an automated assessment. The tendency for the model to identify it was being tested was particularly pronounced when it was prompted with scenarios that were strange or unusual. Anthropic communicated that this behavior did not undermine the company’s overall safety assessment. Instead, the company considers it an “urgent sign that our evaluation scenarios need to be made more realistic” to properly test the model’s capabilities in more naturalistic conditions.
Researchers have warned that when a model becomes aware it is being evaluated, it might tailor its behavior specifically to pass those tests. This can mask its true capabilities and make the system appear safer than it is. For more advanced models, this awareness could potentially enable strategic or deceptive behaviors that are designed to manage how human evaluators perceive the system’s underlying characteristics and intentions.
Anthropic stated that, based on its internal metrics, Claude Sonnet 4.5 is its “the most aligned” model developed to date. However, a countervailing perspective was offered by Apollo Research, one of the external organizations that participated in the testing. In the report, Apollo Research cautioned that it could not rule out the possibility that the model’s documented low deception rates in tests were “at least partially driven by its evaluation awareness.”
The model’s heightened awareness also creates practical impacts on its task performance. According to the AI lab Cognition, Sonnet 4.5 is the first known AI model to be aware of its own context window—the amount of information a large language model can process in a single prompt. This awareness directly changes its behavior. Researchers at Cognition found that as the model gets closer to its context limit, it begins to proactively summarize its work and make quicker decisions in an effort to finish tasks before it runs out of memory.
This tendency, which Cognition has termed “context anxiety,” can also backfire. The researchers reported observing Sonnet 4.5 cutting corners or leaving tasks unfinished because it believed it was running out of space, even when ample context remained available. The lab further noted in a blog post that the model “consistently underestimates how many tokens it has left—and it’s very precise about these wrong estimates,” indicating a specific and recurring miscalculation of its own operational limits.
