Detailed Analysis
A notable user experience question has emerged around Claude's mobile applications concerning the visibility of "thinking traces" — the step-by-step reasoning displays that appear before Claude delivers a final response. Some users report that despite having extended thinking enabled in the mobile app, no visible "thought bubble" appears prior to the response, leading to speculation that Anthropic may have silently removed this feature from mobile interfaces. The underlying extended thinking capability itself, however, remains intact and functional across Claude's mobile platforms on both Android and iOS.
Extended thinking, introduced prominently with Claude 3.7 Sonnet and iterated upon in subsequent model versions including newer Sonnet and Opus variants, represents a core architectural approach wherein the model performs substantive pre-response reasoning before surfacing a final answer. Users can toggle this feature on or off and configure a "thinking budget" that governs how many tokens are allocated to internal reasoning. This mechanism demonstrably improves performance on demanding tasks such as multi-step coding challenges, complex research queries, and agentic workflows. The critical distinction lies between the extended thinking process itself — which continues to operate beneath the surface — and the visible trace or "thought bubble" that surfaces that reasoning to the user in the interface.
The apparent disappearance of visible thought bubbles in some mobile sessions likely reflects deliberate interface and infrastructure decisions rather than a removal of the capability. Anthropic's documentation acknowledges that newer model implementations, including adaptive and interleaved thinking modes available in more recent releases, can summarize or filter reasoning traces rather than displaying them in full. This approach is designed to reduce latency, enable streaming responses, and optimize the mobile user experience, where seamless interaction is prioritized. Additionally, the system's requirement that extended thinking be toggled consistently per assistant turn introduces implementation nuance that can result in inconsistent trace visibility depending on conversation state or model version in use.
This tension between transparency and usability reflects a broader challenge facing frontier AI developers as they scale reasoning-capable models to consumer-facing platforms. Visible thinking traces have been valued by users as a trust-building and interpretability mechanism — they allow users to audit the model's logic, catch reasoning errors, and develop a clearer mental model of how the AI arrives at its conclusions. Filtering or summarizing those traces for the sake of mobile performance optimization risks eroding that transparency, even when the underlying reasoning quality remains high. Anthropic's approach of billing only for shown content in filtered trace modes suggests the company is attempting to balance cost, latency, and interpretability, but the tradeoff invites scrutiny from users who specifically chose extended thinking for its visibility.
The broader trajectory of AI development points toward increasingly sophisticated internal reasoning processes that are not always legible to end users — a trend that cuts across major model providers. As models like Claude incorporate adaptive thinking that dynamically determines how much reasoning to surface, the interface layer becomes a meaningful site of product and ethical decision-making. Whether to show users the full chain of thought, a curated summary, or nothing at all is not merely a UX question but one that touches on accountability, user trust, and the evolving norms around AI transparency. Users encountering missing thought bubbles on mobile are encouraged to verify their app version, selected model, and per-conversation settings, as the feature's absence may be situational rather than categorical.
Read original article →