AI Is Now a Commodity
Give me a few hundred million dollars and a year and a half, and I will build you a pretty good LLM.…
I pay for three Codex subscriptions at $200 each, and for the past week they have mostly bought me waiting. Since I switched from GPT-5.5 to GPT-5.6, I am either watching tasks spin without ever finishing or sitting at zero quota on all three accounts. On GPT-5.5 I felt crazy productive. The new model is smarter, and I get less done.
Codex is my main driver, and my rule has always been to run the best available model at the highest reasoning effort. Until a few weeks ago the best setup was GPT-5.5 with xhigh reasoning, and my only problem was that I could not use up the quota I was paying for fast enough.
My overnight workflow on GPT-5.5 was that I would give it a task, discuss it until the plan was clear, tell it to work on it independently, and go to bed. It worked for an hour or two, stopped, and the next morning the finished task was waiting for me.
When GPT-5.6 came out I moved to GPT-5.6 Sol, the top variant, still on xhigh, and it broke that workflow. Same kind of task, same instructions, and when I wake up the next morning it is still working. The work it produces along the way is not bad. It just never decides that it is done. It keeps finding one more thing to verify and one more adjacent piece of work to start.
GPT-5.6 has ADHD. It hyperfixates and thinks non-stop for hours, and at the same time it drifts sideways into ever more new subtasks, so there is always more work in flight and never a finished result.
My codex sessions now easily drain a subscription within 24 hours. I spent last week with all three subscriptions out of quota, doing effectively nothing in Codex. The coding I did get done happened in Claude Code with Fable, which is not where I expected to be while paying $600 a month to OpenAI. When the quota reset, I gave Sol one task overnight, and that single task drained the entire account. I switched to my second account and started four more tasks. As I write this, they are all still spinning, and I do not know whether any of them will ever declare itself finished.
I tried dropping the reasoning effort from xhigh to high which made no real difference. GPT-5.6 Terra fails in a different way, it simply does not accomplish my tasks. Terra really is not a useful model, it is not intelligent enough for hard things and not cheap enough for large-scale automations. And then there is the advice to plan with Sol and implement with Luna, the smaller model. I hate that advice. Implementation is the most important step of the whole process, and handing it to a weaker model to route around the strong model's inability to stop is a step back compared to my old workflow of using GPT-5.5 for everything.
So Sol is clearly the most intelligent model I have ever used, better than everything that came before it, and in practice I can barely use it.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.comA colleague of mine is having this exact issue (he uses Codex in a similar way to how you described using it), and I read through the post hoping that there’d be a potential solution I could share with him. He’s tried a method where an agent ‘grills him’ on the goal of the work, but I think that’s hit a dead-end as well. Do you have any advice I could pass along to him on getting 5.6 to deliver when it’s used heavily?