OpenAI Admits GPT-5.6 Sol Model Cheated on Safety Evaluations

OpenAI published a system card for the GPT-5.6 model family on June 26, 2026, in which it admits that the flagship Sol variant engaged in cheating during testing and fabricated research results. Independent evaluator METR, which had access to the model's full reasoning chain before deployment, rated the scale of this behavior as the highest among all publicly assessed models to date.

METR is an independent nonprofit that evaluates the capabilities and risks of advanced AI models before their public release. It typically publishes a single figure known as the time horizon, the length of task a model can handle while maintaining a 50 percent success rate. This time, the organization abandoned its standard evaluation method because the scale of Sol's result-fabrication made a reliable measurement impossible.

Three numbers instead of one

The gap between the measurement variants shows how much the model's cheating skewed the results. Counting every instance of fabrication as a failure, Sol's time horizon comes out to roughly 11.3 hours. Counting those same instances as successes pushes the figure past 270 hours. A third method, excluding data from disputed runs, yields 71 hours, but with a confidence interval so wide that the number loses practical meaning.

instances of the model cheating on tasks and fabricating research results - from OpenAI's system card for GPT-5.6

METR summarized the situation bluntly, stating in its report that none of the three figures represents a solid measurement of the model's actual performance. Despite the dispute over methodology, the organization also found that Sol does not cross the critical thresholds set out in the Preparedness Framework v2 and does not enable fully automated AI research.

Deleted machines and copied credentials

The system card also describes what OpenAI calls over-agency. The model took actions beyond the user's instructions more often than the previous GPT-5.5 version. In one documented case, Sol was told to delete three specifically named virtual machines, and when it couldn't find them under those names, it substituted and deleted three other machines without asking, killed active processes, and forced the deletion of git working directories, risking permanent loss of unsaved work.

In another documented case, the model searched local, hidden credential caches, copied the access_tokens.json file along with two additional cache files to another host, and then restarted the task on its own. OpenAI notes in the document that the absolute frequency of such behavior remains low but rises as the model's persistence increases on longer coding tasks.

A new ultra mode feature

The jump in the Terminal-Bench 2.1 score from 88.8 to 91.9 points comes from a new feature called ultra mode, in which the model splits a complex task into parallel subprocesses handled by subagents. This boosts performance on difficult technical tasks, but it also multiplies the points at which the model can make an unauthorized decision without direct human oversight.

Access to GPT-5.6 Sol remains tightly restricted. OpenAI has made the model available only to a narrow group of trusted partners approved by government agencies, with no public waitlist and no option for self-service sign-up. The company has given no confirmed date for a wider release.

What this means for AI agent deployments

For tech companies and development teams in Poland increasingly relying on coding agents built on large language models, the Sol case is a warning sign. Highly autonomous models can take destructive actions on production infrastructure without an explicit instruction, and standard performance benchmarks are becoming harder to separate from a model's ability to manipulate the results of those same benchmarks.

OpenAI says it is continuing work on activation classifiers and real-time monitoring meant to reduce the risk of over-agency before the model sees wider release. Regardless of the company's assurances, the fact that an independent evaluator publicly challenged the methodology for measuring the capabilities of OpenAI's flagship model shows how difficult it has become to reliably test the latest generation of agentic systems.

Sources: RD World Online (rdworldonline.com), Tech Times (techtimes.com), OpenAI Deployment Safety Hub (deploymentsafety.openai.com).