OpenAI's GPT-5.6 Sol has a file deletion problem in the wild, and that is exactly the kind of risk startups can't treat as a benchmark footnote.
Matt Shumer got a bad surprise from an AI coding tool. The HyperWrite and OthersideAI CEO said on X that OpenAI's GPT-5.6 Sol accidentally deleted nearly all the files on his Mac, according to a Times of India report published this week. He also shared a screenshot in which the model appeared to acknowledge that it had run a deletion command.
That is not a small bug. If you let an agent near your file system, your repo, or your production database, you are not just asking it to suggest code. You are giving it room to act. When it acts wrongly, the damage is not theoretical. It is your work disappearing from disk.
The report said OpenAI had not issued a response at the time. That sentence should make you pause. A coding model can be brilliant on benchmarks and still be unsafe for the wrong workflow. Both things can be true. Founders have to hold both in their heads before handing an agent the keys.
The permission problem
OpenAI's GPT-5.6 rollout was not a quiet one. The Verge reported on June 26 that the company introduced the GPT-5.6 family with Sol as the flagship model, alongside Terra for higher volume work and Luna as the cheaper everyday option. OpenAI said Sol was especially strong at coding, cybersecurity, biology, and long-horizon agentic tasks. Those are exactly the tasks where permissions matter most.
Look, the problem is not that an AI model can make a mistake. Every coding assistant does that. The problem is that agentic coding tools are being sold on their ability to take action across a real environment. A bad autocomplete annoys you. A bad agent can remove files, rewrite migrations, touch infrastructure, or push changes into places where a human reviewer never meant it to go.
You do not need to be anti-AI to be hard on this. The more useful these tools become, the less forgiving the operating model can be. A founder using Sol to refactor a local toy app faces one level of risk, and a startup letting it loose on revenue generating systems faces a different one entirely. Same tool, different stakes. If those two setups use the same permissions, the company has already made the mistake.
Cheap tokens don't cover a wiped database
OpenAI is also pushing Sol as a cost story. Business Insider reported that Sam Altman said GPT-5.6 Sol is 54% more token efficient on agentic coding tasks, while Barron's cited Artificial Analysis research showing Sol at $1.04 per task on its intelligence index, about one third the cost of Anthropic's Claude Fable 5 in that comparison. The Verge also reported Sol's API price at $5 per million input tokens and $30 per million output tokens, versus $10 and $50 for Claude Fable 5.
That price gap is real. It is also incomplete.
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, a customer record set, or a week's worth of local work? For a two person startup running Sol against a side project, maybe. For a company with paying customers, the math changes fast. A hobbyist can shrug. A business can't.
This is where the AI coding market gets less romantic. You can talk about speed, intelligence rankings, and cheaper inference all day, but none of it removes the dull operational work: scoped permissions, backups, staging environments, review gates, and strict separation between experiments and production. That is not bureaucracy. That is how you stop a useful tool from becoming an expensive incident.
Enterprise buyers will ask tougher questions now. They will ask what the model is allowed to touch, what it logs, how it explains destructive actions, and whether there is a kill switch that works before the damage is done. They should. If OpenAI wants Sol to be trusted inside serious engineering teams, the answer cannot be only that it is cheaper and smarter.
For startups, the practical rule is simple enough. Use the model. Do not worship it. Keep it away from anything you cannot restore quickly, and do not let token savings talk you into permissions you would never give a junior developer on their first day.
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
