Detailed Analysis
A Reddit user with a Pro subscription reports a frustrating and unusual experience with Claude's Sonnet 4.6 model, in which the AI system exhibited persistent contradictions, circular reasoning, and self-undermining behavior during what should have been a straightforward task: generating a presentation from a document. The user describes a full day of compounding errors, including a notable instance where Claude stated there were nine sections in the document, immediately second-guessed itself, and then declared there were ten — within the same response. Most strikingly, the model reportedly advised the user to abandon the conversation and open a new chat, explicitly acknowledging that its accumulated errors had rendered its own outputs untrustworthy.
The behavior described is consistent with a well-documented phenomenon in large language models known as context window degradation, sometimes called "context poisoning" in informal usage. As a conversation grows longer and accumulates corrections, self-revisions, and contradictory outputs, the model's attention mechanisms must process an increasingly large and internally inconsistent context. This can create a feedback loop where prior errors influence subsequent responses, compounding inaccuracies rather than resolving them. The "Medium effort" setting referenced by the user likely refers to an extended thinking or reasoning budget setting, which can amplify these effects if the model's chain-of-thought reasoning becomes entangled with conflicting prior outputs.
The model's self-referential suggestion to open a new chat is a particularly notable detail. It suggests a degree of metacognitive awareness built into Claude's design — the ability to recognize degraded performance and recommend remediation — but it also exposes a tension in that design philosophy. A system capable of identifying its own unreliability is arguably more trustworthy than one that plows forward confidently, yet the user's experience of watching trust erode in real time across a full day of interaction represents a significant usability failure. The Pro subscription context matters here: users paying for premium access have elevated expectations for reliability and consistency, especially on tasks framed as non-complex.
This incident connects to broader, ongoing challenges in AI development around long-context reliability and stateful conversation management. While frontier models like those in the Claude 3.5 and 3.7 generation have dramatically expanded context windows — sometimes to hundreds of thousands of tokens — raw context length and effective context utilization are not the same thing. Anthropic and its competitors continue to work on maintaining coherence across extended interactions, but production deployments regularly reveal edge cases where performance degrades in ways that controlled benchmarks do not capture. The document-to-presentation task described should theoretically be well within the model's capabilities, suggesting the failure was more about conversational state management than fundamental task difficulty.
What makes this report particularly significant from a product perspective is its novelty to the user — they explicitly note never having experienced this behavior before, implying a regression or an unusual model state rather than a known limitation. This points to potential issues with deployment consistency, where the same model version can behave very differently depending on server-side factors such as load balancing, inference infrastructure, or subtle updates to system prompts and sampling parameters. Anthropic has historically been iterative in its model deployments, rolling out changes incrementally, which means any given user's experience on a specific day may not reflect the model's typical behavior — a transparency gap that can make troubleshooting both for users and for the company itself particularly difficult.
Read original article →