Click any words you think contain an error. Click again to unmark.
Reasoning models often hide or summarize internal chain-of-thought from users — visible CoT is a product choice driven by safety, IP, and UX, not a universal default.
The model may think for pages internally, but the app you use might show only a short answer or a brief summary. Companies hide raw thinking because it can leak unsafe content, reveal trade secrets in how the model reasons, or overwhelm users with messy scratch work. Reasoning happens internally; displaying it is optional.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Reasoning models generate long internal traces — but product APIs decide what end users see. This spot-error crushes a common misconception that reasoning equals visible chain-of-thought in chat.
The flagged span overgeneralizes visibility into an 'always' default.
Generated vs displayed
At inference, reasoning models autoregressively produce thinking tokens — internal chain-of-thought — before emitting the user-facing answer. Those tokens consume GPU time and appear in usage metrics.
Display is a separate layer. The chat client may render only the final completion, a short 'thinking' summary, or nothing at all from the internal channel.
The erroneous claim conflates generation with streaming to end users. Most 2026 frontier products default to hidden or summarized exposure.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- OpenAI o-series generates hidden reasoning tokens billed separately while chat UIs show concise answers or limited summaries.
- Anthropic extended thinking can expose summarized thinking blocks in some clients but not full raw traces to all users.
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy does hiding CoT help safety?
Raw reasoning channels can contain jailbreak intermediate steps users could learn from; hiding reduces attack surface while monitors still scan logs.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Assuming users always see full raw thinking tokens — most frontier APIs hide or summarize traces by default.
60 second bullets to scan on the way to the call.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.