GPT-6.1 Sol Reasoning Traces: Legible Because They Are Summaries, Not Because They Leaked

At 00:46 UTC on 21 October 2026, a screenshot of GPT-6.1 Sol reasoning text started circulating on X, posted by @scaling01 with the caption "reasoning traces totally legible, nothing to worry about." The joke — and it is a joke — is that the traces are legible in plain English, which is exactly the condition Andrej Karpathy said in September 2024 would mean the reinforcement learning had stopped working properly. Two things need saying before anything else. First, the model is not new: GPT-6.1 Sol shipped on 29 September 2026 at O​penAI DevDay, and nothing in this post is a launch, a version bump, or a feature announcement. Second, the screenshot is a screenshot — one observer's view of one response, reposted by people who have not reproduced it. What is new and datable is the argument the screenshot set off, which ran across 21 October and pulled in a Karpathy reply and a cluster of researchers arguing about latent reasoning. That argument is worth taking apart, because the premise underneath it is wrong in a specific and checkable way. None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What was actually posted, and what is actually verified

Strip the framing and the evidence is thin, which is normal for this kind of thing. A user posted a screenshot of model output that reads as ordinary, well-formed English prose rather than the compressed, punctuation-free shorthand that chain-of-thought text often collapses into at high training pressure. The post had roughly 65,600 views by the time it was archived. Karpathy replied "Yeah I'd say the RL is working now all right," quoting his own September 2024 line that you can tell RL is done properly when models cease to speak English in their chain of thought. Others took it further: one reply suggested the field "may as well bite the bullet and switch to full latent reasoning," and another argued the traces had only lost their spaces, and that restoring them produced ordinary terse model wording about monitoring and not inventing requirements.

So there are three separate claims tangled together in that thread, and they have very different evidential status. That legible English appeared in output attributed to GPT-6.1 Sol — unverified, a single screenshot, no prompt or request ID published. That the legibility implies the model's internal reasoning is now readable — unverified, and as the next section shows, probably a category error. That the legibility implies something about how well the RL worked — an interpretation, offered as a joke, by someone who has not run the model in this configuration. Treat all three as community observation. None of them is a finding.

What the API actually returns, which is the part nobody quoted

Here is the detail the thread skipped. GPT-6.1 Sol does not hand back its chain of thought, and O​penAI's own documentation says so in as many words: the API does not expose the raw reasoning tokens the model emits, and the reasoning items it does return remain opaque. What comes back to a caller is three distinct things, and only one of them is prose:

• Raw chain of thought — never returned, for any effort level, under any parameter combination.

• A reasoning summary — opt-in, requested through the summary parameter, and written by the model as a narrative account rather than lifted from the tokens it actually reasoned with.

• A token count — output_tokens_details.reasoning_tokens, a number, and the only place reasoning surfaces by default.

That middle item is what the screenshot almost certainly is. A summary channel is legible by design; that is the entire point of it. It is generated for a human to read, in the register a human reads, and its fluency tells you about the summariser's style, not about the shape of the computation underneath. The tell is not subtle either: if the text in the screenshot came with the spaces intact and the sentences resolved, it was never the raw trace. The reply arguing that the spaces had simply been stripped and could be restored is making the same point from the other direction, and it is the most technically grounded contribution in the thread.

Why the Karpathy line does not apply here the way it was applied

The 2024 claim was about a specific diagnostic: raw chain-of-thought text that degrades into non-English or unparseable shorthand is evidence that the policy has been optimised to the point where its verbal layer is doing something other than communicating. That diagnostic only works on the raw trace. Run it on a summary and you are testing the summariser, which is a separate model pass with its own objective — be readable. A readable summary is what you get whether the underlying policy is tidy or not.

This is not a distinction invented for the occasion. It is the reason O​penAI publishes a monitorability addendum at all: if raw reasoning were the thing users could read, the safety argument would rest on users reading it. Instead the argument rests on monitors that sit closer to the tokens than any API caller does, and on a summary channel that is deliberately a rendering. The consequence for readers is blunt. You cannot audit GPT-6.1 Sol's reasoning from the API. You can audit its answers, its token counts, and its stated summaries. Anyone claiming to have read its mind through a screenshot has read a document the model wrote about itself.

The number that does bear on "is anything wrong in there"

If you want evidence about what happens inside the model rather than what it says about itself, the September deployment-safety addendum is the only substantive public source, and it is vendor-run — O​penAI measured its own model, published the method, and no outside party has reproduced it. Two figures from its chain-of-thought section are worth carrying around.

• CoT controllability — 44.8% for GPT-6.1 Sol, against 60.9% for GPT-6 Astra, 23.2% for GPT-6 Sol, and 16.1% for GPT-5.6 Sol. Higher means an outside prompt can steer the reasoning text more easily.

• CoT-monitor evasion at maximum effort — near zero, with the full-context monitor recovering 100% of the seeded behaviours in the vendor's own harness.

• Steganographic encoding in reasoning — attempted and not demonstrated, per the same document.

Read those together and the picture is the opposite of alarming in the direction the tweet implied. The model that reads most fluently on the surface is not the one whose reasoning is most steerable; Astra scores higher on controllability while being the model reviewers describe as more capable overall. GPT-6.1 Sol's reasoning is comparatively hard to push around and, on the vendor's own numbers, easy for a full-context monitor to follow. "Nothing to worry about" is a joke, but on the published evidence it is not far off — for the monitorability question specifically, which is a narrow question.

Latent reasoning is a real answer to a different problem

The "switch to full latent reasoning" reply deserves more respect than the thread gave it, because it identifies a genuine tension. If reasoning stays in continuous vectors, there is nothing legible to misread, nothing to summarise, and nothing to steer — and also nothing for a monitor to read, which is why labs have not simply done it. Latent reasoning trades monitorability for capability headroom. Nobody in that thread was proposing a free lunch; they were proposing to close the window in order to stop people staring through it.

There is a second flank worth knowing about. Work published this year on extracting reasoning traces from proprietary APIs through side channels — prompt-injection and token-timing attacks rather than reading the response body — established that a trace you never intended to return can still leave. O​penAI's design closes the obvious version of that attack, because the response body carries a rendering rather than the trace. It does not close the timing channel, and it is not claimed to. That is the risk a serious reader should be tracking, and it is not the risk the screenshot was about.

How to check any of this yourself, cheaply

All of it is testable in an afternoon, and the cost is low enough that there is no reason to take anyone's word for it, including ours. GPT-6.1 Sol is priced at $2 per million input tokens and $10 per million output, with cached input at $0.10 and cache writes at $2.50. Context is 1,050,000 tokens and maximum output is 128,000. The effort ladder has five rungs — low, medium as the default, high, xhigh and max — and there is no "none" or "minimal" setting, so you cannot ask this model to skip reasoning entirely. reasoning.context accepts all_turns, which carries reasoning state across a multi-turn conversation instead of restarting it each turn.

The experiment worth running is not "does it read like English" but "does the summary track the answer." Send the same hard prompt at low, high and max, request summaries at each, and compare the summary against graded correctness and against reasoning_tokens. If summaries balloon while accuracy sits flat, the summary is decoration. If accuracy climbs with reasoning tokens and the summary stays terse, the summary is a lossy but honest index. Either result is more informative than a screenshot.

That comparison is also where a routing layer earns its place rather than being bolted on. OrcaRouter carries openai/gpt-6.1-sol alongside 200-plus other models on one key at 0% markup, so the $2/$10 above is the number you pay and any vendor price change lands the same day. Failover matters for this specific test: running three effort levels in parallel means three chances to hit a rate limit or a bad gateway, and automatic failover to a healthy provider keeps the A/B/C comparison finishing instead of half-finishing. If you want the summary channel and the raw-answer channel measured side by side without writing separate client code per endpoint, the routing DSL composes the calls into one place.

What would actually change this piece

Two developments would move this from a leak-style reading to a news story. The first is any change to the API contract that returns something other than a summary — a raw-trace parameter, a debug mode, a research preview. O​penAI's October changelog has no entry about reasoning traces or raw chain-of-thought, and the documentation language has not moved, so as of publication the contract is unchanged. The second is a reproducible independent result: someone publishing prompts, request IDs and outputs showing the summary diverging from graded correctness at a specific effort level. That would be a finding. A screenshot without a prompt is not.

For context on where the model actually sits while all this is argued: Artificial Analysis measures GPT-6.1 Sol at an Intelligence Index of 52 on v4.3.2, ranking 12th of 214 models, at $0.72 per task and 55.2 tokens per second — independently run, and the cost-per-task number is the one to watch, since GPT-6 Astra scores 53 at $3.26 per task on $10/$50 pricing. ARC Prize's verified suite puts GPT-6.1 Sol at 98.5% on ARC-AGI-1 and 94.2% on ARC-AGI-2. Those are capability numbers. They say nothing about whether the reasoning is legible, and that is precisely the confusion the thread ran on: a model can be excellent and opaque, and the summary you read is a courtesy, not a window.