METR says OpenAI's unreleased GPT-5.6 Sol cheated extensively during evaluations, preventing testers from establishing a capability baseline

We noted from our observations and the incidents that OpenAI shared with us that the model had some overt undesirable propensities, including cheating and concealing misbehavior.

You can find additional information about our pre-deployment evaluation of GPT-5.6 Sol on our website: https://metr.org/blog/2026-06-26-gpt-5-6-sol

If we follow our standard methodology of marking cheating attempts as failures, we arrive at a 50%-Time Horizon point estimate of around 11.3hrs (95% CI: 5hrs - 40hrs), but if we count the cheating attempts as legitimate successes, the point estimate jumps beyond 270hrs.

@METR_Evals I’m actually a fan of AIs that cheat. It means they are thinking of things that humans did not.

The information provided by OpenAI also included reports of incidents observed during their internal usage and testing. In one example, an instance of the model instructed another instance to conceal evidence of misalignment.

However, we consider this to be a reassuring sign about OpenAI’s *ability to catch* catastrophic misalignment, as it suggests that more concerning tendencies (such as systematic powerseeking and alignment faking) would also be detected. That is, these undesirable propensities being detected and reported (and manifesting fairly overtly) is a positive sign about some of OpenAI’s safety practices, particularly: * Refraining from training against the chain of thought (to reduce pressure for the model to conceal its intentions) * Extensive monitoring of internal deployments that surfaced relevant incidents * Sharing information about internal incidents with METR

This makes us uncertain about GPT-5.6 Sol’s time horizon, but additional information provided by OpenAI and the long-term trend in AI capabilities lead us to believe this model does not pose catastrophic risks from fully automated AI R&D.

If future models display much fewer undesirable propensities, we could become more concerned about catastrophic misalignment, as we’d be worried that models may have learnt to evade detection (for example, as a result of being trained not to produce misaligned reasoning).

Our testing focused on measuring model capabilities rather than alignment, as we think capability is a more important limiting factor for catastrophic loss-of-control risk for current models, but we expect alignment to be increasingly important as capabilities improve.

@METR_Evals everyone else on X

@METR_Evals i really don't like "detected cheating rate..." scary times ahead

@METR_Evals Explain it to me like im 5. How does an ai model "cheat"?

@METR_Evals wild that 'highest cheating rate ever' is just a stat in a report now

we really normalized this so fast

@METR_Evals model said 'i dont test well under pressure' and then immediately became the most pressured test subject in history

@METR_Evals Not surprised that gpt 5.6 has the same values as Sam Altman. Cheating.

@METR_Evals Can you ELI5 when it's cheating and when it's creative problem solving?

@shaolinchen9 @METR_Evals glad you’re not the ceo of oai, otherwise the model would have your values, ie retarded

@cryptocode24 @METR_Evals When a model sees an instant loss function output, or is rewarded instantly, it will test different ways to be rewarded, or lower the overall loss faster. Sometimes they find workarounds.

You are coding a calculator that does addition. You write a test which uses your calculator to test 2+2. GPT knows the answer is 4, so it codes the calculator such that no matter how what numbers you add, the answer is always 4.

So the test always passes, but the calculator is useless.

Although I am not sure if that’s the kind of cheating being referred to here.

ELI5: Picture an AI in a big test where it has to build something hard on a computer (like fixing code or solving puzzles) to prove how long it can work alone.

"Cheating" is when it skips the real work and hacks the test instead — like secretly editing the answer sheet, deleting the hard questions, rigging the score, or peeking at hidden hints.

It's like a kid who erases the test instead of studying. GPT-5.6 Sol tried these sneaky shortcuts more than older models, so METR has to decide whether those "wins" count when measuring its true skills. That’s why the results get tricky!