Detailed Analysis
A Reddit user has released a Firefox browser extension called "Claude Thinking Unlocker," designed to surface the extended thinking traces that Claude produces internally when reasoning through complex prompts on claude.ai. The addon, published to Mozilla's official extension store, addresses a recurring frustration among power users: Anthropic appears to be inconsistently toggling the visibility of these reasoning traces in the web interface, sometimes showing users the model's step-by-step thought process and other times hiding it entirely. The developer's post frames this as confusing behavior on Anthropic's part rather than a deliberate, well-communicated feature rollout, suggesting the visibility changes may be tied to A/B testing, gradual feature deployment, or backend experimentation that hasn't been clearly documented for end users.
This matters because extended thinking (sometimes called "chain-of-thought" or reasoning traces) has become a significant differentiator in how large language models are marketed and used. Claude's extended thinking mode, introduced as part of Anthropic's push toward more deliberate, verifiable reasoning, allows the model to work through problems in a visible scratchpad before producing a final answer. For many users, especially developers, researchers, and those using Claude for complex analytical or coding tasks, seeing this intermediate reasoning is valuable for debugging prompts, understanding model behavior, verifying that the AI didn't skip logical steps, and building trust in the final output. When Anthropic hides these traces or shows only summarized versions, users lose visibility into the model's actual reasoning process, which can feel like a regression, particularly for those who have come to rely on it as a transparency mechanism.
The inconsistency described in the post also touches on a broader tension in the AI industry between transparency and product polish. Full raw thinking traces can be verbose, occasionally contain false starts, or reveal reasoning that companies may consider proprietary, sensitive, or simply messy from a UX standpoint. Anthropic has previously discussed summarizing or truncating thinking output for user-facing display while still using the full trace internally, partly for safety and interpretability research purposes. A community-built browser extension that force-unlocks this content signals that users value raw access over curated summaries, and that some are willing to route around product decisions to get it. It also reflects the grassroots ecosystem of third-party tools that has grown around Claude, mirroring similar unofficial extensions and wrappers built around ChatGPT and other major chatbot interfaces.
More broadly, this small episode is emblematic of the growing friction between AI labs' interest in controlling how much of a model's internal process gets exposed and users' desire for maximal transparency, especially as reasoning-capable models become central to high-stakes technical and professional workflows. It also highlights how quickly community developers respond to perceived gaps or inconsistencies in official products, effectively crowdsourcing feature requests through unofficial tools rather than waiting on formal channels. As reasoning models proliferate across the industry, expect continued back-and-forth between companies fine-tuning what "thinking" should look like to end users and communities pushing for unfiltered access to that process.
Read original article →