[FEATURE] Provide unsummarised, raw thinking to interactive clients

Problem Statement

We want the unsummarised, raw thinking. Not a summary of it.

What renders in an interactive session is a separate model's paraphrase. Measured across 38,726 thinking turns in my own transcripts, the ratio of visible thinking text to billed thinking_tokens:

| model | n | median ratio |
|---|---|---|
| opus-5 | 20,047 | 0.40 |
| sonnet-5 | 11,453 | 0.36 |
| fable-5 | 7,201 | 0.35 |

| billed thinking tokens | n | median ratio |
|---|---|---|
| 0–200 | 11,841 | 0.55 |
| 200–500 | 12,785 | 0.37 |
| 500–1,000 | 8,532 | 0.31 |
| 1,000–2,000 | 4,135 | 0.28 |
| 2,000+ | 1,433 | 0.28 |

Every frontier model compresses at the same rate, and the ratio halves as reasoning grows. A long thought shows roughly a quarter of itself — compression is heaviest exactly where there is most reasoning to inspect.

The thinking is generated. It is billed. It is not returned.

(One user's own transcripts, so the workload is mine rather than representative. Visible tokens estimated as chars/4, applied identically across buckets, so the trend holds whatever the estimator.)

Proposed Solution

Return unsummarised thinking to interactive clients, gated. A setting behind a flag, a paid tier, an enterprise agreement, or an explicit opt-in — any of those works.

To be explicit, because the ask could be read the other way: I am not asking for raw thinking on by default. If every subscription returned raw traces, harvesting them at scale would become trivial, which is precisely the exposure the summarisation appears to exist to prevent. A gate is not a compromise I am settling for — it is what I am asking for.

Legible English is preferable, but is not a requirement and should not be a reason to withhold it. If the raw trace is fragmentary, compressed, repetitive, non-English, or not recognisably natural language at all, return it as it is. Readability is the user's problem to solve, not a bar the output has to clear first — and if models reason in something other than clean prose, that is worth seeing rather than a reason to hide it.

If the answer is no, say so in the docs and say why. A documented "no" closes this completely and would be accepted here without argument.

What does not work is the current state, where the distinction is not stated anywhere and users reasonably believe they are seeing reasoning.

Alternative Solutions

showThinkingSummaries: true is the closest available and does not do this — it appears to switch between summaries and redacted placeholders rather than between summaries and raw. The ratios above were measured with it enabled. If that reading is wrong, documenting the actual raw path resolves this issue on its own.

There is no workaround.

Use Case Example

1. An agent runs unattended and takes a wrong turn at step 40.
2. The summary for that turn reads plausibly and does not contain the wrong turn.
3. Billed thinking for that turn is roughly four times what is visible.
4. The reasoning that would explain the failure was generated and paid for, and cannot be read.

Additional Context

I assume you have reasons, and that they are better informed than mine — you have the extraction data and I do not. If the answer is distillation risk, that is a good answer and the compression curve above supports it rather than undermining it.

Two things I would rather say than leave implied. Raw traces make reasoning auditable by anyone, including where it is unflattering — that cost falls on you, not on me, and I think it is worth carrying but would not pretend it is free. And the faithfulness question (arXiv:2505.05410) is the reason this is not just convenience: an unlabelled paraphrase is harder to check than a labelled one, and arXiv:2608.09867's Figure 8 is the worked example.

Its closing line is why I am asking rather than dropping it: "an architectural design that hides a user's own data from them - yet leaves it entirely vulnerable to third-party extraction - provides neither privacy nor security."

Related: #76104 (original request, closed), #90115 (narrower ask: mark summarised blocks with a metadata field).

Reproducing the measurement is about twenty lines against any transcript directory — please check it rather than take my word for it.

💡 written with Consciousness 😉