Newer frontier reasoning models have substantially more controllable chains of thought than earlier generations.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The strongest evidence comes from OpenAI. Its 2026 system cards report that GPT-6-generation models score far higher on the CoT-Control evaluation than GPT-5.x models when chain-of-thought length is held fixed. For chains of 750 to 1,250 tokens, GPT-6 Astra controls 60.9% of its chains of thought, GPT-6.1 Sol 44.8% and GPT-6 Sol 23.2%, against 16.1% for GPT-5.6 Sol and 1.7% for GPT-5.5. OpenAI treats this as a regression in monitorability. It also reports that Astra's controllability rose during RL training, whereas earlier models' controllability did not. Within OpenAI's newest generation the jump is large, and it holds with length held fixed.
As a general statement about frontier models, the claim is weaker. The trend has not been steady even at OpenAI: GPT-5.5 was less controllable than GPT-5.4 and GPT-5.2. At Anthropic, controllability appears to track model size rather than release date. Claude Opus 5.5 is reported as less controllable than the larger Claude Mythos 5.1. The CoT-Control paper found that newer Claude variants resisted repeated-attempt elicitation better than older ones. All the figures come from developers measuring their own models, and none has been independently replicated.
The claim is therefore well supported for OpenAI's latest generation, but not yet shown as a trend across developers. Comparable length-matched CoT-Control results from other labs' newest models, or an independent run of the open-source suite on current frontier models, would settle which reading holds.
Full reasoning: the evidence and decisions behind this verdict
Prior: the March 2026 CoT-Control paper (arXiv 2603.05706) found low controllability across 13 models (0.1% to 15.4%). It found controllability rising with parameter count and falling with more RL and longer reasoning. I expected generational changes to depend on the training recipe rather than run steadily upward.
Evidence: snippets from the OpenAI Deployment Safety Hub (the GPT-6 Astra card, published 2026-09-03, its CoT-controllability page, and the GPT-6.1 Sol addendum) and the Astra safety overview. Together these confirm that the GPT-6 models are "meaningfully higher" than GPT-5.6 Sol, with length controlled. The figures come from the parent claim's earlier reading of the same cards. The GPT-5.5 card says GPT-5.5's controllability is lower than GPT-5.4 Thinking and GPT-5.2 Thinking, so the trend is not monotone. A LessWrong summary of the Claude Opus 5.5 system card (Sept 2026) says Opus 5.5 has lower controllability than Claude Mythos 5.1 and matches other Opus models, below the higher Mythos-size rate. I read this from a secondary summary and did not check the card's own figures. The CoT-Control paper notes that newer Claude variants (e.g. Sonnet 4.5) resist repeated attempts better.
Weighing: the claim's most natural reading is a broad generational trend. The direct evidence for it comes from one developer's latest step. That step is large (roughly 1.5x to 4x at matched length) and backed by a training-dynamics change, which counts strongly for that developer. The counterevidence shows the pattern is not universal or monotone. I chose supported over inconclusive because the only frontier generation clearly measured on a matched-length basis shows a substantial rise. Confidence is modest because Anthropic's data cut against generality. Would change: length-matched CoT-Control results from Anthropic, Google or others on their newest models, or independent replication.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because OpenAI's GPT-6-generation models score several times higher on CoT-Control than GPT-5.x models with chain-of-thought length held fixed, and GPT-6 Astra's controllability rose during RL training, unlike earlier models, the newest frontier generation is taken to have substantially more controllable chains of thought than its predecessors.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 supported1 not separately assessed
If its premises hold, the argument shows a substantial generational rise for OpenAI's models, because the comparison controls for chain-of-thought length. It establishes the general claim only to the extent that one developer's latest generation stands for frontier models as a whole. The argument rests on GPT-6-generation OpenAI models score substantially higher than GPT-5.x models on CoT-Control at matched chain-of-thought length., which the system cards report directly but which no one has replicated independently. GPT-6 Astra's chain-of-thought controllability increased during RL training, unlike earlier OpenAI models. adds corroboration at the level of mechanism.
Because among Claude models controllability tracks model size rather than release date, with the newest Opus model less controllable than its predecessor, and because OpenAI's own GPT-5.5 was less controllable than GPT-5.4 and GPT-5.2, newer frontier models are not in general more controllable; the rise is confined to one developer's latest generation.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 unknown
If its premises hold, the argument shows the trend is not universal or monotone. It does not undo the large jump in OpenAI's latest generation, so it bears mainly on how general the claim can be. It rests on Among Claude models, chain-of-thought controllability tracks model size rather than release date., which comes from secondary summaries of Anthropic system cards and has not yet been assessed.
Provenance
Where this claim has been said, linked to its canonical form.
We find that GPT-6 Sol and GPT-6.1 Sol's CoT controllability is meaningfully higher than that of GPT-5.6 Sol.
System card section on CoT controllability, reporting scores by CoT length.
We have found that GPT‑6 Astra is more capable of controlling its own CoT than GPT‑5.6 Sol, and less likely to include incriminating information in its CoT.
Safety overview summarizing monitorability regressions in GPT-6 Astra.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.