Context Anxiety

I mentioned the following finding to a friend recently but hadn't posted about it here, so rectifying that.

Cognition Team on 2025-09-29:

Sonnet 4.5 is the first model we've seen that is aware of its own context window, and this shapes how it behaves. As it approaches context limits, we've observed it proactively summarizing its progress and becoming more decisive about implementing fixes to close out tasks.

This "context anxiety" can actually hurt performance: we found the model taking shortcuts or leaving tasks incomplete when it believed it was near the end of its window, even when it had plenty of room left.

I did some digging and found only the Haiku and Sonnet models receive an injected context awareness statement behind the API. That means every harness using them is subject to this issue. It's interesting to me that they consider the feature's effectiveness acceptable on Sonnet and Haiku, but chose not to apply it to Opus despite releasing new versions multiple times this year.

I remember this unintuitive result whenever I'm thinking about the prompting system for an LLM workflow or agent — to be careful of feeding "useful" information without having a system for testing results.