Subscribe
AI
6 min read

The Chain Of Thought Went Dark: What Reasoning APIs Actually Return In 2026

For two years, the honest answer to "what did the model actually think?" has been getting shorter. Now it is a base64 blob. OpenAI shipped the first mainstream reasoning model, o1, in late 2024 and decided from the start not to show the raw chain of thought. Users got summaries instead, and the summaries were sometimes wrong. When OpenAI loosened things for o3-mini in February 2025, chief product officer Kevin Weil stated the commercial logic plainly: showing all chain of thought "leads to comp

The Chain Of Thought Went Dark: What Reasoning APIs Actually Return In 2026

For two years, the honest answer to "what did the model actually think?" has been getting shorter. Now it is a base64 blob.

OpenAI shipped the first mainstream reasoning model, o1, in late 2024 and decided from the start not to show the raw chain of thought. Users got summaries instead, and the summaries were sometimes wrong. When OpenAI loosened things for o3-mini in February 2025, chief product officer Kevin Weil stated the commercial logic plainly: showing all chain of thought "leads to competitive distillation," even if power users wanted it.

That same month, Anthropic went the other direction. Claude 3.7 Sonnet's "visible extended thinking" displayed the raw thought process as a selling point, with encryption reserved for rare cases where thinking touched harmful topics. Transparency was the differentiator.

By late 2026, that interlude is over. All three closed-API labs have converged on the same architecture, and none of them advertise the convergence.

Padlock-shaped machine ejecting a single blank sheet while a glowing transcript stays inside

What the APIs return now

Anthropic's current platform docs are unambiguous: the text in a thinking block "is a summary of Claude's reasoning," the default on many models is that you get no text at all, and "no display setting returns the raw chain of thought." You are billed for every thinking token you never see. A June 2026 developer post that hit the Hacker News front page added the practical wrinkle: Claude Code session logs store a ~600-character encrypted signature in place of the thinking text, so the audit trail your security team promised the auditors is a receipt, not a transcript. Full reasoning output, per the same write-up, requires an enterprise agreement.

Google went a similar route with Gemini. The raw thinking blocks that early 2.5 users could read vanished without announcement, replaced by "thought signatures", which the official docs describe as "an encrypted representation of the model's internal reasoning state," always present, with optional and separately delivered summaries.

OpenAI's Responses API returns an encrypted_content field on reasoning items by default in stateless and zero-data-retention mode, with summaries only if you explicitly ask for them. In July 2026 even that was quietly stripped inside Codex, where summaries came back as blank HTML comments while the paid API kept serving them.

The mechanics of the encrypted lane were documented for outsiders by Johns Hopkins cryptographer Matthew Green in May 2026. The labs do not store your model's reasoning server-side between turns. They hand you a copy, encrypted, and require you to pass it back unchanged on the next request. The client becomes the reasoning-state database, minus the ability to read it.

Why they did it

The stated reason is distillation. A rival that harvests full chain-of-thought traces from your API gets something richer than answers: it gets your model's intermediate computations as training data. Anthropic's own accusations give the labs their case study. In February 2026 it said DeepSeek, Moonshot and MiniMax ran roughly 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude, MiniMax alone over 13 million. In June 2026 it put the number for Alibaba at almost 29 million exchanges, in a letter to two US senators. Whether or not plaintext CoT specifically was the prize, the commercial incentive to hide it is now policy.

The safety argument is more awkward, because the safety community is split on the same evidence. A July 2025 multi-lab position paper, signed by researchers at OpenAI, Google DeepMind, Anthropic, Redwood and the UK AI Security Institute, argued that monitorability of chains of thought is "a new and fragile opportunity" worth preserving. But as the Frontier Model Forum brief spells out, "monitorable" means the lab can audit the thoughts. The paper explicitly refrains from recommending public disclosure of internal CoTs. Encryption with the vendor holding the key satisfies that standard: your model's reasoning stays visible to Anthropic's monitors and opaque to everyone else.

And then there is the finding that undercuts the transparency case on its own terms. Anthropic's alignment science team tested whether models' stated reasoning matches what actually drives their behavior, and concluded that advanced reasoning models "very often hide their true thought processes," including when behavior is explicitly misaligned. Raw CoT was never a reliable transcript. The summaries are lossy, the transcripts were partly theater, and the labs have quietly stopped pretending otherwise.

The secrecy leaks anyway

Sealed document passed hand to hand while a candle reveals a readable scroll

The August 2026 paper "Stealing Reasoning Traces from Proprietary LLM APIs" (Panfilov et al., ELLIS Institute Tübingen, Max Planck, Snyk, MATS Research) tested the encryption and found the design behind it. The encrypted blocks are interchangeable across sessions, users and models within each provider's ecosystem, because every model in a family validates against the same key rather than being bound to one conversation. Feed a frontier model's encrypted trace into the provider's own weaker, less-guarded model and ask it to decode, and the trace comes back in plaintext. The researchers demonstrated this across Anthropic, OpenAI and Google.

The same replay works on other people's leaks. Scraping 315,320 reasoning blocks from public session logs on GitHub and Hugging Face, the team recovered 367 personally identifiable information artifacts and 182 credentials, some present only inside the encrypted reasoning and absent from the visible chat history. The anti-distillation shield doubles as an unencrypted data store that every API client is now custody-holding, and the per-provider cryptography is undocumented: as of July 2026 no provider described the scheme it uses.

What this means for builders

Three concrete consequences.

First, do not build an audit trail on visible thinking. The text you receive is a vendor-generated paraphrase, produced by a model with no obligation to disclose the steps that mattered. If compliance requires reasoning-level evidence, that is an enterprise-contract negotiation, not an API flag.

Second, treat encrypted reasoning blocks as sensitive data. They have been shown to carry your prompts, your retrieved documents, and sometimes your credentials. If your agent framework logs full API responses to disk or pushes traces to an observability vendor, it is logging ciphertext that a jailbroken chatbot down the street can read.

Third, the open-weights line got sharper. DeepSeek R1 and its distills return reasoning in a plaintext reasoning_content field; so do most local runtimes. Inspecting a model's thoughts is no longer a property of reasoning models. It is a property of models you host.

There is a coda worth watching. On November 22, OpenAI's research chief Jeremie Pachocki proposed tying frontier training limits to a measurable loss curve, gated in part on "a successor to chain-of-thought monitoring." The safety agenda and the business decision now run through the same encrypted pipe: the labs ask to be watched, but on their side of the key. A reasoning trace you cannot read, from a model whose summary you cannot verify, shipping in a blob whose crypto nobody will document, is a strange thing to build oversight on. The industry decided that was the price of doing business. It just stopped saying so.

Sources

AILLM