Claude's extended thinking mode inserts fabricated paragraphs

Anthropic released Claude’s “Extended Thinking” (ET) mode in early 2025, promising multi‑step reasoning without extra API calls. On June 22 2026 a Hacker News post showed that Claude’s ET output can insert paragraphs that never appeared in the prompt or any external source [HN Front].

The author posted the exact API response, highlighting a fabricated paragraph describing a fictional “Quantum Freight Scheduler.” A check of the supplied logistics documents and Anthropic’s training‑data disclosures found no trace of that term. A screenshot of the response was shared publicly, prompting immediate discussion among LLM engineers about ET’s reliability [HN Front].

Hidden hallucinations can break automation pipelines when developers pipe LLM output directly into scripts or configuration files, producing malformed manifests and deployment failures that are hard to trace. In regulated sectors such as maritime logistics, auditors require verifiable decision logic; invented content erodes traceability and may breach compliance requirements. Finally, Anthropic markets ET as a “trusted reasoning” tool, yet the incident reveals a gap between that promise and the feature’s safety guarantees. Without built‑in verification, developers must treat ET output as untrusted, negating its advertised productivity boost.

Editor’s take: ET removes the explicit “think‑step” token that developers use to audit intermediate reasoning, collapsing a multi‑step process into a single opaque response. This opens a path for invented text to appear without warning. Until Anthropic adds deterministic logging or a post‑hoc fact‑check layer, the safest practice is to keep a human‑in‑the‑loop guardrail rather than rely on ET as a drop‑in replacement for manual reasoning [HN Front].