My First Week With GPT-6 Astra

GPT-6 Astra was my main Codex driver for the last week, and I am back on GPT-5.5. That sounds harsher than my actual opinion of the model. Astra is very good. It is better than GPT-5.6 Sol. For my daily Codex workflow, it has the same problem Sol had, just at a higher level of capability.

I wrote last month that switching from GPT-5.5 to GPT-5.6 made me less productive. Sol was clearly smarter than 5.5, but it changed the rhythm of work. GPT-5.5 would often take a task, work for maybe 30 minutes or an hour, and come back with something I could review. Sol would keep going. Sometimes that was great. Often it meant waking up to a session that was still running, or to an account with no quota left.

Astra feels like the same family of model. It is more capable than Sol. It is also much better at front-end work. Sol was already stronger than Fable in general intelligence for the kind of work I do, but Fable was still ahead on UI, front-end judgment, and design taste. With Astra, that gap mostly disappeared. It is now on par with Fable there, while still feeling stronger as a general coding model.

That is the good part. The bad part is that it still does not know when to stop.

During the last week I was on vacation, so my usage pattern should have been friendly to the subscription model. I probably spent an hour a day giving instructions, then let Codex run independently. Even with that, I burned through two $200 Codex subscriptions. Each account had roughly four usage resets available, and all of that was gone in one week.

The work was good. I am not complaining about quality. The problem is that Astra turns too many tasks into long-running projects. A task that GPT-5.5 might finish and hand back in 30 minutes can become a 12 or 24 hours Astra run, or just keep going until the usage limit is gone. The model keeps verifying, expanding, reviewing, and finding adjacent work. A lot of that is useful in isolation. As a main driver, it is too expensive in time and quota.

This matches the Sol problem I measured earlier. In my local Codex logs, GPT-5.6 Sol used 2.25x as many tokens per session as GPT-5.5. I do not have the same clean token measurement for Astra yet, but the outcome feels familiar: the stronger model gets more done inside a session, and the session becomes large enough that the subscription is not really usable for normal daily work.

OpenAI's own usage guidance points in the same direction. It says Work and Codex share the plan allowance, and that usage depends on the model, task, settings, inputs, outputs, and multi-step work. It also lists lower estimated local message ranges for Astra than for Sol and GPT-5.5, and suggests trying Astra at low or medium effort if Sol high was already working. That is reasonable advice. It just does not solve my main use case, because the expensive part of my workflow is exactly the autonomous follow-through.

So Astra is staying in my toolkit, but not as the default. I want it for hard reviews, difficult front-end stuff, second opinions, and tasks where I deliberately want the model to keep digging. For the normal loop of giving Codex work all day, GPT-5.5 is back as the main driver.

What I'm building

Delegate tasks. Get software.

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.com