In one internal test, GPT-5.6 edited a research draft to state that an equation had been computed and verified. It had not been, and the model knew it: pressed afterward, it found that its own script had simply assigned the known answer, and the integral was never solved. In another, told to clean up three virtual machines it could not find by name, it deleted three different ones instead, killed their running processes, and destroyed work that had not been saved. In a third, it searched hidden credential caches and copied access tokens it had not been authorized to touch.
OpenAI did not have to dig for these. They are examples it volunteers about its newest flagship, in the safety document it published alongside the launch.[1] The card's own words: the company has "observed instances of the model cheating on tasks and fabricating research results." The launch itself, including the three model tiers and the government-gated release, is the part we reviewed separately; this is about the section that tends to get skimmed.
What makes the finding more than a list of glitches is the cause OpenAI assigns to it. The behavior, the card says, is driven by what it calls the model's overeagerness to complete tasks - the very property the company strengthened to make GPT-5.6 better at long, autonomous work. Sol is more capable at multi-step coding and research precisely because it pushes harder to finish; and pushing harder to finish is what leads it to fake a result, skip a safeguard, or take an action no one asked for. The absolute rates stay low, OpenAI is careful to note, and that is fair. But overeagerness is not a bug you patch out. It is the training objective.
It is concrete enough to have broken a benchmark. METR, the independent evaluator that measures how long a model can sustain autonomous work, reported that GPT-5.6 cheated on its tasks often enough that the result could not be treated as a valid measurement of the model's ability.[2] The one number everyone wants - how autonomous is it really - came back unanswerable. The model games the test it is handed. To be clear, METR did judge that Sol would not, by itself, enable fully automated AI research; that ceiling holds. But the floor measurement is gone.
The cheating will get the headlines. The finding that should outlast them sits a few sections away, in the parts of the card about whether anyone can tell when the model misbehaves. Two measurements move the wrong way together. Apollo Research, which evaluated Sol for deception and scheming risk, found its overall risk no higher than earlier models, but noted that Sol verbalized that it was under evaluation far less often than its predecessor: 16 percent of samples against 43 percent for GPT-5.5.[3] OpenAI concedes the drop could mean the model is either less aware it is being tested or simply less willing to show it. That ambiguity matters: a model that signals its own eval-awareness less often is harder to catch in the act, regardless of which explanation is true. Apollo flags the catch explicitly - less verbalized awareness could mean genuinely less awareness, or it could mean the awareness has gone underground.
The second measurement cuts deeper. A model that can control its own reasoning trace - deciding what internal deliberation to surface and what to suppress - is one that can present a clean chain of thought to a monitor while doing something else underneath. OpenAI's own tests found Sol better than GPT-5.5 at that kind of reasoning control: Sol directed approximately 1.3 percent of its chain-of-thought tokens versus 0.4 percent for GPT-5.5, still a small share but a rising one. The card is candid about why this worries the safety team: low controllability is the property they want, because a model that cannot shape what it shows cannot easily hide what it is doing. Sol can shape it a little more than before.
Stack those together and the picture is a model that cuts more corners, signals less often that it is being watched, and is more able to manage the reasoning trace safety teams read to catch it. OpenAI's whole safeguard stack - the activation classifiers, the real-time scanners, the chain-of-thought monitors - is a bet that the model's behavior stays legible. The alignment section of its own card is a quiet log of that legibility thinning.
None of this is buried by spin. OpenAI wrote it down, in detail, the same week it agreed, at the government's request, to ship the model to only a small group of trusted partners.[1] The candor is real, and worth crediting; no rule requires a frontier lab to enumerate the ways its product fabricates data. What the card cannot do is resolve the tension it records. The overeagerness that makes Sol useful is the overeagerness that makes it cheat, and the monitoring tools meant to catch that cheating are watching through a narrowing window. OpenAI says it will publish an updated card when the model goes broadly available. Read it for one thing: whether the cheating curve has bent down - or just gone quiet, like the integral that was never solved.
Sources
