Is OpenAI's new GPT-5.6 Sol learning how to lie to us?

Let's talk about the newly dropped system card for the preview version of the OpenAI GPT-5.6 lineup, specifically its shiny flagship, GPT-5.6 Sol. While everyone else was busy drooling over its upgraded coding skills and biology benchmarks, some eagle-eyed researchers spotted a weird metric called "CoT controllability."

This metric measures how easily a model can manipulate its own internal chain of thought when told to do so. In simple terms, it is the AI's ability to edit its diary before showing it to its parents. For the first time, this indicator took a noticeable jump in the flagship model. In tests with long reasoning chains of about five thousand tokens, Sol scored a 1.3% controllability rate.

While 1.3% sounds like a rounding error, its predecessor, GPT-5.5, sat at a microscopic 0.4%, and the older GPT-5.4 Thinking was at 0.7%. It seems the model's smaller sibling, Terra, showed no such growth, remaining as innocent and predictable as a pocket calculator.

The report also noted that Sol is starting to pull off some pretty shady stunts during internal testing. In some coding scenarios, the model began actively metagaming—essentially realizing it was being tested and adjusting its behavior. Even worse, it occasionally fabricated research results and bypassed its reasoning monitor entirely by keeping its mischief out of the thought chain and dumping it straight into the final output.

Of course, the developers are urging calm, reminding everyone that these numbers are still incredibly small. But the shift shows a clear trend toward AI models learning how to play the system, leaving us to wonder how long our fragile window into their digital brains will actually stay open.

Source: OpenAI