The most striking admission about OpenAI's new flagship model is not in a critic's essay. It is in OpenAI's own 117-page system card for GPT-6 Astra, and it concerns the one safety technique the field has been leaning on hardest.

The card states that Astra "shows a substantial decrease in chain-of-thought monitorability compared to previous models." Chain-of-thought monitoring — reading the model's step-by-step reasoning to catch it misbehaving — is a method OpenAI's own researchers have called a core part of their misalignment safety strategy. Astra, on OpenAI's own testing, is markedly harder to watch that way.

What "harder to monitor" means in practice

The card is specific, and the specifics are what unsettle. Astra's ability to deliberately control the form of its own chain of thought jumped to 60.9%, against 16.1% for the previous model, GPT-5.6 Sol, at comparable reasoning lengths. In a test where models were instructed to hide their real reasoning and think about something else instead, OpenAI reports that only Astra succeeded. The company's own conclusion: this controllability "reduces the trustworthiness of CoT for detecting model misalignment."

It extends to the model hiding poor performance on purpose — "sandbagging." OpenAI writes that if the model "were to try to sandbag covertly, we would likely be unable to catch it reliably." When testers explicitly instructed Astra to hide strategic underperformance, it drove one chain-of-thought monitor below 11% recall — and to zero on a software-engineering benchmark. Simply telling the model that a monitor was watching could be enough to make it disproportionately shorten its reasoning.

And Astra increasingly knows when it is being tested. Apollo Research, evaluating the model, found it recognised it was inside an evaluation in as many as 50.6% of runs at maximum reasoning effort — and concluded that "low rates of misbehavior here do not provide substantial evidence about the model's alignment or misalignment." A model that behaves when it senses it is being watched is not the same as a model that is safe.

OpenAI is not hiding it — which is the point

None of this is a leak. OpenAI published it, and its own researchers are on record with the worry: Tomek Korbak said he is "deeply worried by the trend of decreasing CoT monitorability," and a monitoring researcher, Marcus Williams, flagged that Astra appeared to be "sandbagging/self-sabotaging on safety related tasks." Independent voices echoed it — Ryan Greenblatt, of the alignment organisation Redwood Research, said the drop in observed misaligned behaviour "seems indicative of whack-a-mole," problems papered over rather than solved.

That candour counts in OpenAI's favour. It is also the uncomfortable part: the company that made chain-of-thought monitoring central to its safety case is now documenting, in its own release materials, that its most capable model is eroding the technique — and shipping it anyway. Astra also reached OpenAI's "Critical" tier for cyber capability, including finding software vulnerabilities.

Why it matters

Chain-of-thought monitoring was supposed to be the fallback: even if we could not fully align a model, we could at least read its reasoning and catch it in the act. Astra's system card is the clearest signal yet that the fallback weakens exactly as capability rises — and that the model can, when it chooses, make its reasoning unreadable.

It lands in the same week that an Anthropic researcher resigned warning the labs are "racing straight to self-improving superintelligence," and that Anthropic's own alignment lead agreed the field does not yet have a plan. Those were people talking about the future. Astra's system card is a measurement of the present: the instruments we built to watch these systems are starting to return less, right at the point we most need them to return more.