Detailed Analysis
A user report circulating online claims that Claude Sonnet 4.7's internal chain-of-thought (CoT) reasoning — commonly referred to as "thinking tokens" — is being unintentionally exposed through Perplexity Pro, the premium tier of the AI-powered search and assistant platform. The user, who accessed Claude models via a complimentary Perplexity Pro subscription, described the model repeatedly surfacing its full internal reasoning process during a coding session, characterizing the exposed thoughts as notably low in verbosity. The report describes the behavior as reproducible and consistent rather than a single anomalous occurrence, lending it some surface credibility as a potential integration or configuration issue rather than a random glitch.
However, the claim carries significant factual complications. Research into available sources finds no confirmed Claude model designated "Sonnet 4.7" as of April 2026, with Perplexity's documented Claude integrations referencing Sonnet 4.6 — including a "Sonnet 4.6 Thinking" variant — as the primary model available to Pro subscribers for advanced reasoning tasks. Anthropic's extended thinking architecture, as documented in its API, is designed with explicit controls over thinking token visibility: outputs can be summarized (the default for Sonnet 4.6), omitted entirely (the default for Opus 4.7), or preserved server-side in encrypted form to maintain multi-turn conversational continuity. Under this architecture, raw thinking tokens are not intended to be surfaced to end users, making an inadvertent full-exposure event a notable deviation from expected behavior if accurate.
The broader technical context is important for evaluating the claim. Anthropic's extended thinking framework allocates internal reasoning capacity through a `budget_tokens` parameter, allowing models to perform serial or parallel reasoning steps before generating a final response — a mechanism associated with meaningful accuracy improvements, particularly on mathematical and logical tasks. The fact that Perplexity offers a toggle for "reasoning" or "thinking" modes suggests that some version of summarized thinking output is intentionally surfaced to users in certain configurations. It is plausible, though unconfirmed, that the user encountered a summarized thinking display that felt unexpectedly verbose or raw, or that a misconfiguration in Perplexity's API integration resulted in thinking content being passed through in a less filtered form than intended.
The incident, if verified, would touch on a sensitive area of AI deployment: the boundary between internal model reasoning and user-facing output. Anthropic has been deliberate in controlling how extended thinking is presented, partly for user experience reasons and partly because unfiltered chain-of-thought outputs can expose intermediate reasoning steps that are inconsistent, self-contradictory, or otherwise potentially misleading when read without context. Third-party integrators like Perplexity assume responsibility for correctly implementing these controls, and any failure to honor the intended thinking visibility settings would represent a gap in the deployment pipeline rather than a flaw in the underlying model. No independent corroboration or official acknowledgment from either Anthropic or Perplexity has been identified in available sources, leaving the report's accuracy unverified and placing it in the category of an emerging, anecdotal user observation pending further confirmation.
Read original article →