Ask HN: Dear Anthropic, can we please have thought traces back?

An Ask HN thread titled “Dear Anthropic, can we please have thought traces back?” is live on Hacker News with 5 points and 6 comments. The poster asks Anthropic to restore thought traces for Claude, arguing that without them it is hard to tell whether the model reached an answer by valid reasoning, by cheating on the task setup, or by fudging and inventing intermediate steps. They single out Opus 4.6 as still the best model in their view precisely because of how its reasoning can be inspected when those traces are available.

Thought traces are the intermediate reasoning steps a model exposes before the final answer. With them, a user can walk the chain of claims, spot jumps, check that intermediate facts match the prompt, and catch fabricated citations or silent assumptions. Without them, evaluation collapses to the final string: correct-looking output can hide shortcut behavior, prompt leakage, or confabulated steps that never appear in the UI. The HN ask is therefore about observability of the model’s path, not about a new capability or benchmark score.

For engineers and builders, that observability is a practical debugging and trust tool. Teams that use models for coding, research, or agent loops need to know whether a correct answer came from a sound plan or from lucky pattern match and invented detail. Missing traces force slower, more expensive verification: re-running prompts, cross-checking sources by hand, or building external logging that the product no longer provides. The poster’s complaint maps directly onto day-to-day QA of LLM-assisted work.

On the competitive side, the thread frames Opus 4.6 as still preferred when thought traces are present, even as the same ecosystem is discussed alongside Codex and Fable on HN. That is a product-positioning signal: transparency of intermediate reasoning is treated as a differentiator, not a nice-to-have. Vendors that hide or strip traces may win on polish or latency but lose users who need to audit how conclusions were formed.

What to watch is whether Anthropic restores thought traces in the product surface the poster is using, and whether similar asks spread for other models that omit intermediate steps. Until then, anyone evaluating Claude—or comparing it to alternatives—should treat final-answer-only output as unverifiable on method, and design review workflows that do not assume the model’s path is visible.