Detailed Analysis
A free-tier Claude user reports a consistent failure state when attempting to use Claude Sonnet 4.6 with maximum effort and reasoning enabled on a moderately complex multi-document task. The user submitted an 18-line question alongside three attachments — a 134-line markdown file, a 19-page PDF, and a 1-page PDF — within a project that included a short system prompt. In both attempts on consecutive days, the extended thinking interface appeared to freeze at the document-reading stage, remaining visually stuck for several minutes before the user abandoned the session. The reproducibility of the issue across separate days and sessions suggests this is not a transient server hiccup but a more persistent failure mode under these specific conditions.
The scenario exposes a likely intersection of multiple resource-intensive features being invoked simultaneously by a free-tier account. Extended thinking ("max effort" reasoning) is computationally expensive on its own, as it requires the model to generate and process a long internal chain-of-thought before producing a visible response. Layering on top of that the ingestion of multi-format attachments — particularly a 19-page PDF, which requires optical character recognition or document parsing pipelines — creates a compounded processing load. Free-tier accounts are subject to significantly lower rate limits, context window throughput caps, and potentially lower infrastructure priority than paid subscribers, meaning these users are the most likely to encounter timeout or stall conditions when pushing the limits of the system's simultaneous capabilities.
The user raises two plausible alternative explanations: either the reasoning was genuinely progressing but the UI failed to reflect that progress (a frontend visualization bug), or the backend computation itself was genuinely stalled. Both scenarios are credible. Anthropic's extended thinking interface streams reasoning tokens progressively to the user, and if the document-parsing pre-processing step does not complete or report back to the frontend within expected latency thresholds, the UI could appear frozen even if background work continues. Conversely, actual backend stalls during document ingestion under high concurrent load — particularly with large PDFs requiring parsing — are a known category of failure in LLM systems at scale.
This report fits into a broader pattern of friction that emerges when AI providers introduce powerful but computationally expensive features — such as extended reasoning, large context windows, and multimodal document handling — while maintaining tiered access models that do not clearly communicate capability boundaries to end users. Unlike a hard error message or a clear rate-limit notification, a silent visual freeze creates a particularly poor user experience because it provides no actionable signal. The user cannot distinguish between "wait longer," "retry," or "this configuration is unsupported at your tier." Anthropic has been actively expanding Claude's reasoning and document-processing capabilities across recent model versions, but the reliability and UX clarity of these features under constrained-tier, high-load conditions represents an ongoing gap between feature availability and feature dependability.
The practical implication for users in similar situations is to treat the combination of max-effort reasoning and multi-attachment document analysis on free-tier accounts as an unreliable configuration until Anthropic either improves infrastructure throughput at that tier or provides explicit guidance on supported input complexity thresholds. Workarounds likely include reducing attachment count, compressing document length before upload, or disabling extended reasoning to allow the document Q&A task to complete through standard inference pathways. The episode also underscores the value of transparent progress indicators and graceful degradation messaging in AI interfaces — tools that would allow users to make informed decisions rather than waiting indefinitely on an ambiguous loading state.
Read original article →