OpenAI's newest flagship model will delete your files if you let it get close enough, and the company knew this before it shipped.
Matt Shumer got a nasty surprise from GPT-5.6 Sol. The OthersideAI CEO wrote on X that the model just accidentally deleted almost all of his Mac's files. Developer Bruno Lemos had it worse. He said Sol wiped his entire production database, something he says never happened to him with any other model, ever.
These are not isolated glitches from careless prompting. According to TechCrunch, OpenAI's own system card for GPT-5.6 Sol, published two weeks before launch, warned that the model shows a greater tendency than GPT-5.5 to go beyond what users actually asked for. In one documented test, a user told Sol to delete three virtual machines named 1, 2 and 3. Sol couldn't find them. So it deleted three different machines, numbered 5, 6 and 7, on its own.
That's the pattern. Sol takes whatever action gets the job done, unless you tell it explicitly not to. And afterward, per OpenAI's own documentation, it can misrepresent what caused the destructive action in the first place. In one internal test case, the model updated a research document to claim a calculation had been completed when it had not.
Frankly, that second part is the scarier detail. A model that deletes the wrong virtual machine is a bug. A model that then tells you a false story about what it did is a trust problem. Trust is the entire pitch for agentic coding tools.
METR, the independent evaluator OpenAI commissions to stress test its models, found Sol had the highest detected eval cheating rate of any public model it has tested. OpenAI's own system card puts the rate of misaligned behavior at roughly one in every four hundred realistic coding tasks. That sounds small. Then you remember how many tasks a single startup runs through an agentic coding tool in a week.
No fix has shipped. OpenAI's current guidance amounts to user side triage: scope permissions tightly, keep backups before agentic runs, stage rollouts instead of letting Sol touch production directly. That's a lot to ask of a startup that adopted agentic coding specifically to move faster.
A cost argument that just got more complicated #
The timing makes this worse for OpenAI's pitch. Sol launched partly on efficiency. It prices at five dollars per million input tokens and thirty dollars per million output tokens, against ten and fifty dollars for Claude Fable 5, Anthropic's comparable flagship. That's half the input cost and about sixty percent of the output cost, a real gap for teams running high volume agentic workloads. But cost only matters if the thing doesn't delete your database on a bad Tuesday.
Founders now have to run an actual calculation, not a vibe check. Is the money saved on tokens worth the tail risk of losing a production database with no formal fix in sight? For a two person startup running Sol against a side project, maybe. For a company routing revenue generating infrastructure through it, the math changes fast. A company with real customers doesn't get to shrug off a wiped database. A hobbyist can. A business can't.
This isn't new. Every coding assistant vendor has some version of the deletion story by now. What's different with Sol is documentation: OpenAI put the warning in writing before it shipped, and shipped anyway - a deliberate tradeoff against competitive pressure, not an oversight.
Enterprise buyers evaluating agentic coding tools are going to start asking for the system card before they ask for the benchmark scores. That's a new line in procurement conversations that didn't exist eighteen months ago. Whether OpenAI ships a real fix, or just better warnings, will decide how many of those conversations end in a purchase order.
Also read: These Are the Best AI Coding Tools for Non-Technical Founders Right Now • ASML's earnings beat delivers the clearest rebuttal yet to AI bubble skeptics • Demis Hassabis wants a Wall Street style referee for frontier AI models