GPT-6 Astra's chain-of-thought controllability increased during RL training, unlike earlier OpenAI models.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The GPT-6 Astra system card, published by OpenAI on 3 September 2026, states this directly: "Unlike previous models, we find that CoT controllability for Astra especially increases over the course of RL training." OpenAI says it is investigating the cause and whether the increase can be reversed in future training runs. Chain-of-thought controllability is a model's ability to follow instructions about the content of its own reasoning, measured with the CoT-Control suite of more than 13,000 tasks.
The card documents the end state in detail. Among chains of thought 750 to 1,250 tokens long, Astra complied with the control instructions 60.9% of the time, against 16.1% for GPT-5.6 Sol and 1.7% for GPT-5.5 Thinking. The pattern held across three datasets and eight instruction types. Much less is shown for the specific claim about the training trajectory. It is stated rather than plotted in the sections read, and the comparison with how earlier models behaved during their own RL training rests on OpenAI's internal checkpoint measurements, which no outside party has checked.
The finding runs against the developer's interest, because it weakens the case for relying on chain-of-thought monitoring. That makes a misreport unlikely. The claim is therefore well supported as an account of OpenAI's measurements. It would reach verified if checkpoint-level data or an independent replication were published.
Full reasoning: the evidence and decisions behind this verdict
The primary source was read directly: the GPT-6 Astra system card at deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf and its web version, section 9.2.1 "CoT Controllability". The web text gives the CoT-Control method (Yueh-Han, 2026) and reports controllability scores conditioned on CoT length: 60.9% for Astra, 16.1% for GPT-5.6 Sol and 1.7% for GPT-5.5 Thinking, at 750 to 1,250 tokens. These results are consistent across datasets and instruction types. The card's own text names the training-trajectory claim, in a passage found through search in both the PDF and the web card: "Unlike previous models, we find that CoT controllability for Astra especially increases over the course of RL training." The same passage says OpenAI is confident there are no significant direct CoT grading pressures, and that it is trying to understand whether the increases can be reverted.
The claim's two components therefore have the developer's explicit statement behind them: (1) controllability rose during Astra's RL training, and (2) this differs from earlier OpenAI models. The evidence is first-party, though, and the per-checkpoint trajectory and the earlier-model baselines were not seen in the sections read. That is why the status is supported rather than verified. No contrary statement or independent evaluation was found. The search that failed to find one was brief: the exact phrase plus the system card. External evaluators such as UK AISI appear in the card, but no evidence was found that they measured the training trajectory.
What would change the verdict: published checkpoint curves or an outside replication would move it up. A system-card correction, or a showing that earlier models were not measured comparably during RL, would move it down.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
We are actively investigating the causes of GPT-6 Astra's increased CoT controllability. Unlike previous models, we find that CoT controllability for Astra especially increases over the course of RL training.
Monitorability section. OpenAI discusses why Astra's CoT controllability on the CoT-Control suite is higher than that of GPT-5.6 Sol and GPT-5.5, and what it is doing to understand and possibly reverse the increase.
Assessment history
0 status changes over 2 assessments. full history →
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.