GPT-6 Sol first test: an 11-point list of problems
The author lists 11 issues including tunnel vision, messy code, tweaking tests to fit bugs, and constant regressions, calling Sol the worst frontier model he has used, even worse than 5.6.
First impressions of GPT-6 Sol I've run it on several of my projects, and I already have plenty to say. «GPT. GPT never changes» 1. Tunnel vision and getting endlessly stuck on a single task? Check ✔ 2. Disgusting code that turns your whole project into a mess? Check ✔ 3. Zero understanding of what tests are or how to write them, plus tweaking the tests to fit its own bugs? Check ✔ 4. Can't deliver a result on the first try without errors? Check ✔ 5. Every new change introduces a pile of regressions? Check ✔ 6. Text that's painful to read? Check. 7. The intelligence of a boiled shrimp? Check ✔ 8. Zero emotional intelligence and a poor grasp of intent? Check ✔ 9. Overengineering and needless complexity? Check ✔ 10. Can't break its work into logical multi-step chains? Check ✔ 11. Can't anticipate consequences? Check ✔ Conclusion GPT-6 Sol is the same classic GPT model with exactly the same problems, except this time they somehow managed to make it even worse. It's GPT-5.5 Mini, GPT-5.6-Flash Sol, GPT-5.6-Light. Call it whatever you want, but it's definitely not GPT-6. It's the worst frontier model I've ever used. It's bad at everything. It's even worse than 5.6 For some reason, OpenAI decided that a lower price would be enough to win people over. It's painfully dumb and gets stuck on tasks like no model before it. You constantly have to ask it to redo things. You waste a ton of time, iterations, and energy. In the end, it costs more than Opus 5.5 I work at a large corporation with 5,000 people. My employer pays for our Claude, Codex, Qwen, and Kimi subscriptions. I don't know a single engineer at the company who hasn't run into every problem I listed. People simply don't enjoy working with GPT models. They don't want to trust them with their code or with serious tasks. Until OpenAI recognizes and fixes this, they'll never earn recognition in the professional community. Their models will stay stuck doing Computer Use and other routine grunt work