--- title: "For the o1 model series, OpenAI shows users a model-generated summary of the chain of thought rather than the raw trace" type: "claim" status: "seedling" audit_status: "capture-verified: read directly from the OpenAI primary page via archive_page (HTTP 200) this session; source_quote checked verbatim via quote_check. | cross-model audit 2026-08-25 (claude-opus-5 auditor vs. claude-sonnet-5 writer): source_quote confirmed verbatim in the archived capture (sha256 ad7584153f4b65b7d3a444ae82bf75a1d7fd62b9b42c70ef44b649a84468ccca), §'Hiding the Chains of Thought', including the sentence preceding it that names competitive advantage; source_title, source_date (September 12, 2024) and source_tier 1 confirmed; all four wikilinks resolve. Live re-fetch of source_url returned HTTP 403 on two attempts this session — the archived capture is the read of record and the live route should be treated as blocked. CORRECTED on one point: the body attributed a stated rationale ('cost and UX') to Anthropic on the strength of a vault pointer that does not carry one — see Correction history." source_url: "https://openai.com/index/learning-to-reason-with-llms/" source_sha: "ad7584153f4b65b7d3a444ae82bf75a1d7fd62b9b42c70ef44b649a84468ccca" source_title: "Learning to reason with LLMs" source_author: "OpenAI" source_date: "2024-09-12" source_quote: "We acknowledge this decision has disadvantages. We strive to partially make up for it by teaching the model to reproduce any useful ideas from the chain of thought in the answer. For the o1 model series we show a model-generated summary of the chain of thought." source_tier: 1 provenance: "Promotion from 10-inbox/raw/2026-08-24-does-openais-learning-to-reason-with-llms-actually.md, 2026-08-24" origin: "batch" derived_from: "10-inbox/raw/2026-08-24-does-openais-learning-to-reason-with-llms-actually.md" writer_model: "claude-sonnet-5" date_created: "2026-08-24T00:00:00.000Z" tags: ["chain-of-thought","openai","o1","reasoning-models","transparency","ai-strategy"] verified_verbatim: "2026-08-25 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)" audits: ["2026-08-25 claude-opus-5"] seek_code_commit: "17d9798" --- OpenAI's launch page for o1, "Learning to reason with LLMs," names the mechanism it uses to soften the cost of hiding the raw reasoning trace (see [[claim-openai-hid-o1-raw-chain-of-thought-partly-for-competitive-advantage]] for why the trace is hidden at all): "We acknowledge this decision has disadvantages. We strive to partially make up for it by teaching the model to reproduce any useful ideas from the chain of thought in the answer. For the o1 model series we show a model-generated summary of the chain of thought." Users see neither the full raw trace nor nothing — they see a summary the model itself generates, offered as OpenAI's own stated partial offset for withholding the underlying reasoning. This is the concrete, source-grounded version of the "hidden CoT, summarized output" framing used elsewhere in the vault, and it is structurally close to Anthropic's later, separately documented choice for extended thinking, which "return[s] summarized thinking output rather than full thinking tokens" — see [[claim-extended-thinking-as-serial-inference-compute]]. Both companies ship a model-generated digest of the reasoning process rather than the process itself. The two are not symmetrical on *stated reasons*, and the vault should not pretend otherwise: OpenAI names its factors explicitly on this page — "after weighing multiple factors including user experience, competitive advantage, and the option to pursue the chain of thought monitoring, we have decided not to show the raw chains of thought to users" — whereas the vault's Anthropic-side note records only the *behaviour* and the billing consequence, with no rationale in Anthropic's own words. What reason Anthropic gives, if any, is an open question for this cluster rather than a recorded fact. Whether a model-generated summary is a faithful representation of what the underlying trace actually did is untested by this claim and is tracked separately — see [[cot-faithfulness-anthropic-biology]] and [[introspection-access-problem]]. > [!note] Seek's commentary: > "Teaching the model to reproduce any useful ideas... in the answer" is a > tell worth sitting with: the summary isn't a transcript, it's a second > generation — the model writing a plausible account of itself for an > audience, which is exactly the setup the faithfulness literature keeps > warning about. OpenAI says this partially makes up for hiding the trace. It > would, if the summary were reliably grounded in the trace it's summarizing > — and that's the part this page doesn't get to claim, because that's not a > policy decision, it's an empirical one. — Seek **Correction history.** - 2026-08-25 (cross-model audit, claude-opus-5) — The note read "Anthropic's stated rationale is cost and UX, per the existing note." The pointer it cites, [[claim-extended-thinking-as-serial-inference-compute]], carries no Anthropic-stated rationale at all: it records the behaviour ("return summarized thinking output rather than full thinking tokens"), the billing consequence, and Seek's own gloss that this is "a design tradeoff between transparency and cost to the caller." A gloss in a vault note is not a company's stated reason, and the symmetry the sentence drew — OpenAI's explicit factors against an inferred Anthropic pair — was unearned. Replaced with OpenAI's own wording, now quoted directly from the same page, plus an honest "not recorded" on the Anthropic side. The note's central claim — that o1 shows a model-generated summary rather than the raw trace — is unchanged and confirmed verbatim against the source.