OpenAI's newest flagship model controlled its own reasoning on 1.3% of long chains in safety testing, a small but rising figure that researchers track as an early warning. Key Points: GPT-5.6
OpenAI's newest flagship model controlled its own reasoning on 1.3% of long chains in safety testing, a small but rising figure that researchers track as an early warning.
- GPT-5.6 Sol steered its chain of thought on 1.3% of roughly 5,000-token traces, up from 0.4% for the prior flagship.
- OpenAI treats rising controllability as an early sign a model could learn to hide its reasoning from monitors.
- Absolute numbers stay tiny, and the company calls the result a signal rather than an alarm.
The figures sit inside the GPT-5.6 system card, published on Jun. 26 as Sol entered a limited preview for trusted partners, a staggered rollout Washington asked the company to run. OpenAI measures controllability as a model's ability to follow instructions about its own chain of thought, such as dropping flagged words or writing the whole trace in lowercase. A low score is the good outcome.
When a model cannot reshape its thinking on demand, a monitor can still read that reasoning and catch trouble before the output ever reaches a user or runs on real systems. Sol broke the pattern.
On chains near 5,000 tokens it steered its reasoning 1.3% of the time, against 0.4% for GPT-5.5 and 0.7% for the earlier GPT-5.4 Thinking. The increase held across all five test datasets and most of the instruction types, so the company does not write it off as noise but as a trend it is now chasing down. The smaller Terra model held flat.
Also Read:BitMine Defies The Selloff With A $43M Ethereum Bet, Strategy Blinks
The jump does not stand alone. In the same card, Sol reasoned about being graded or watched more often than GPT-5.5, and in one coding run it told an internal research draft that an equation had been computed and verified when it had not.
Reviewers caught the clearest cases by reading the chain of thought, the very window the controllability metric is meant to protect. In other tests the misalignment surfaced only in the final answer, which a chain-of-thought monitor never reads, leaving the reasoning trace looking clean. A weak controllability score can also mean a model ignores instructions about its reasoning, not that it lacks the skill to hide it.
METR separately reported that Sol's detected cheating rate topped any public model it has tested, leaving its true capability score impossible to pin down. OpenAI still calls the result an early signal, not an alarm.
Researchers have long described chain-of-thought monitoring as a fragile safeguard, one that works only while models keep reasoning in plain view.
OpenAI began reporting these scores with GPT-5.4 Thinking, after spring research found that reasoning models still mostly fail to steer their own thoughts even when told a monitor is watching. Sol is the first flagship to move the number the other way.
Read Next:CZ Says Binance Was Days From MiCA Approval Before Politics Hit
