GPT-6 Astra's chain-of-thought controllability gains track its no-chain-of-thought capability gains.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The claim is that GPT-6 Astra's rise in chain-of-thought controllability, its ability to follow instructions about the content and form of its own reasoning, moves together with its rise in capability when reasoning is switched off. Its direct basis is OpenAI's own testimony. The Astra system card says the controllability changes "are correlated with the increase in no-CoT capabilities", and OpenAI researcher Tomek Korbak adds that the correlation is strong "across a few generations of models". Neither source publishes the data, the correlation coefficient, the number of checkpoints or models involved, or how no-CoT capability was measured.
The parts of the picture that can be checked publicly agree with the claim. Both quantities jumped sharply at Astra. Astra can complete far more tasks without chain of thought than earlier models: the UK AI Security Institute puts its no-CoT task horizon at roughly 31 minutes, against about 3.6 for GPT-5.6 Sol. Length-matched controllability rose from 16.1% for GPT-5.6 Sol to 60.9% for Astra, on chains of thought of 750 to 1,250 tokens. Controllability also rose during Astra's RL training, the stage at which reasoning-free capability would plausibly be acquired. Both measures also look anomalous for Astra relative to earlier trends, which is the pattern tracking predicts.
The evidence falls short of verification for three reasons. "Tracking" across a few model generations rests on only a handful of data points. The analysis is the developer's own and is unreplicated. And at least one public observation needs explaining: GPT-5.5 was less controllable than GPT-5.4 and GPT-5.2, so the claim requires its no-CoT capability to have dipped as well, which has not been shown. Even if the correlation is real, it does not settle the cause. Capability, RL regime, or architecture could each produce both trends, and that causal question belongs to the parent claim, not this one. Publishing the paired controllability and no-CoT scores across checkpoints and models, or an independent run of the open CoT-Control suite alongside a no-CoT benchmark, would settle the claim.
Full reasoning: the evidence and decisions behind this verdict
Prior model before reading: if no-CoT capability and controllability both reflect a model's ability to do latent computation, they should co-vary. Correlation claims made by a developer about its own models are usually true as descriptions of its internal data, but they are often thinly based. I expected support at moderate confidence, and the reading did not move that much.
Evidence read: - GPT-6 Astra system card (deploymentsafety.openai.com/gpt-6-astra), section 9 overview. It says controllability "especially increases over the course of RL training", unlike for earlier models. It says OpenAI is fairly confident there is no significant direct CoT grading pressure, and that the changes "are correlated with the increase in no-CoT capabilities". No figure or statistic is given. Section 9.2.1 gives length-matched CoT-Control figures: 60.9% for Astra against 16.1% for GPT-5.6 Sol and 1.7% for GPT-5.5 Thinking, on chains of thought of 750 to 1,250 tokens. The same overview says the set of tasks Astra can do with reasoning disabled is "greatly expanded" and cites the UK AISI's roughly tenfold no-CoT horizon increase. - Korbak thread (x.com/tomekkorbak/status/2095596848581071020, Sep 3, 2026): "strongly correlated with no-CoT capabilities across a few generations of models". Root-causing is ongoing. - Zvi Mowshowitz (thezvi.substack.com/p/astra-is-hard-to-monitor) quotes the thread. His disagreement concerns whether capability explains the monitorability decline. He does not dispute the reported correlation. - Related graph assessments. The newer-models controllability claim lists further figures: GPT-6 Sol 23.2% and GPT-6.1 Sol 44.8%, and GPT-5.5 below GPT-5.4 and GPT-5.2. The no-CoT horizon excess claim is supported at 0.5: about 30.9 minutes for Astra against 3.6 minutes for Sol, with an informal expectation of 5 to 9 minutes.
Weighing: no source denies the claim. The publicly observable co-movement is consistent with it, since both measures jump at Astra and both rise during RL, but with so few models even unrelated trends that both increase with scale would co-move. The weight rests on the developer's internal analysis, which is credible as a description but unpublished. The GPT-5.5 non-monotonicity is a small unexplained point. Credence 0.75, matching the parent Steward's seed, reached independently. The status is supported rather than verified because the evidence is testimony, not data.
Would change the verdict: published paired series across checkpoints and models (upward to verified, or down if the relation is loose); evidence that GPT-5.5's no-CoT capability did not dip; or an 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.
OpenAI, which holds the training-run and evaluation data, reports in the GPT-6 Astra system card that controllability changes "are correlated with the increase in no-CoT capabilities", and an OpenAI researcher adds that the correlation is strong across a few model generations. The public record shows both series moving together where it can be checked: Astra can complete far more tasks without chain of thought than earlier models, GPT-6-generation models score far higher than GPT-5.x models on length-matched controllability, and Astra's controllability rose during RL training, the period over which no-CoT capability is also acquired. On the premise that the developer's internal analysis is competent and honestly reported, the two gains track each other.
- 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?
- 2 supported1 not separately assessed
Suppose the developer's internal analysis is competent and honestly reported, and the public co-movement holds. Then the two gains very probably track each other. It is not certain, because "correlated" across a few models or checkpoints can be loose, and the statistic is not published. Of the public premises, Astra's large no-CoT capability gain and the length-matched controllability jump are well attested in the system card and external evaluation. The rise during RL adds the within-training dimension. The weight falls on the testimony premise, which no independent party has checked.
Provenance
Where this claim has been said, linked to its canonical form.
We have found that these changes are correlated with the increase in no-CoT capabilities (i.e., being able to take actions or return an answer without any verbalized reasoning required)
Monitorability overview, CoT controllability paragraph: OpenAI says it is investigating Astra's increased controllability, which rose over RL training, and reports the correlation with no-CoT capability without showing the data.
So far, we found that CoT controllability has been increasing over the course of RL (which wasn't the case for previous models) and is strongly correlated with no-CoT capabilities across a few generations of models.
Thread accompanying the Astra system card describing ongoing root-causing of the controllability increase.
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.