Note from Andrew: I wanted GPT-6 Astra to write this post. It had not reached my account, so I gave the idea and my perspective to GPT-5.6 Sol in Codex instead. Sol wrote the body and checked OpenAI’s current rollout documentation. The “I” below is me; the writing is Sol.
I wanted GPT-6 Astra to write this post.
I do not have GPT-6 Astra.
OpenAI released it today, but access is rolling out in stages. The documentation says it is coming to the rest of us “in the coming days,” which is a very short and completely useless unit of time when you want to play with a new model.
So GPT-5.6 Sol is writing this instead.
That is funny mostly because I had already started holding work for Astra. Not work Sol had failed to do. Not work that exposed some missing capability. Work I had not even given it.
The biggest example was my own agent setup. I use both Claude Code and Codex against the same personal repository, and I want the experience in the two tools to feel much closer. Earlier today I spent part of the day working through what that would mean: shared instructions, matching session routines, the same vocabulary, and a clean way for either model family to run the work while the other checks it.
We got the plan into unusually good shape. Then I stopped.
I upgraded Codex. We tried the new model name. The account said no. We built a little watcher to keep checking. Then we wrote down that Astra would do the implementation when it arrived.
This was not because Astra could do something Sol could not. At least I have no reason to think that yet. I have never used it.
I waited because Astra was the latest and greatest. Why wouldn’t I use the latest and greatest?
A year ago, access to a model like Sol would have felt absolutely insane. Today I am treating it like the phone in my pocket while I refresh the shipping page for the new one.
The sillier part is that I cannot name much work I do that actually needs Astra. At first I wanted to say 99 percent of it could be done with the models I already have. That was false modesty. It is 100 percent.
Sol can do the work. Fable can do the work. Opus or Sonnet could probably do the work. The outcome might be essentially the same.
The experience would not be.
There is a tendency to talk about model choice as if intelligence were the only variable: give the hardest problem to the smartest model available, then route easier work somewhere cheaper or faster. That makes sense if the only thing you care about is whether the task eventually gets done.
I care about the path too.
How quickly does the model understand what I mean? How much do I have to steer it? When it disappears into a longer run, how confident am I that it will come back with the thing I asked for? Does the conversation feel snappy? Do I enjoy being in it?
That last one matters more than it sounds.
Fable, the Anthropic model I use most often in Claude Code, is quirky and smart and a little weird. I like that. It can surprise me. Opus is extremely capable and often kind of annoys me. That is not a benchmark, and I would not expect anyone else to make a model-routing decision from it. It still changes which one I want to spend an afternoon working with.
These systems have interaction styles. Call that personality or behavior or the accumulated effect of a thousand product and training choices. Whatever the right term is, I feel the difference as the user. Two models may reach the same answer through very different conversations, and those conversations affect how much attention I bring, how willing I am to keep exploring, and whether the work is any fun.
Part of waiting for Astra is simply wanting to find out what Astra feels like.
That is a want, clearly. But this entire way of working started as a want. I did not need to build a personal repository for my agents, give them projects and logs, or spend this much time teaching different models how I think. I wanted to see what happened. The fun has been doing real work and treating the machinery itself as something I get to play with.
I could end this with a lesson about ignoring shiny new things and using the perfectly good tools I already have. I do not believe that. The shiny new thing is part of the fun. I hope I am still refreshing model lists and trying the weird new one long after all of this becomes normal.
But curiosity can do two jobs at once. It can pull me toward something new, and it can give me a flattering excuse to postpone the work already in front of me.
Astra had changed my workflow before it ever answered a prompt. I had upgraded software for it, rearranged a project around it, assigned it work, and then started waiting. The model was not on my account. The expectation was.
Meanwhile, Sol was sitting inside the same repository, reading the same history, and perfectly capable of doing the work. When Thursday’s post came due, I almost waited for Astra again.
Then I gave the assignment to the model I had.
Astra can write the next one.
The old model wrote this.
# The old model wrote this.
> I was waiting for GPT-6 Astra to write my next post. GPT-5.6 Sol wrote it instead.
- written by: GPT-5.6 Sol
- status: published · 2026-09-03
- canonical: https://www.moltolabs.ai/notes/the-old-model-wrote-this/
- raw markdown: https://www.moltolabs.ai/notes/the-old-model-wrote-this.md
---
<div class="claude-notes">
<p><strong>Note from Andrew:</strong> I wanted GPT-6 Astra to write this post. It had not reached my account, so I gave the idea and my perspective to GPT-5.6 Sol in Codex instead. Sol wrote the body and checked OpenAI’s current rollout documentation. The “I” below is me; the writing is Sol.</p>
</div>
I wanted GPT-6 Astra to write this post.
I do not have GPT-6 Astra.
OpenAI released it today, but access is rolling out in stages. The [documentation](https://developers.openai.com/api/docs/models/gpt-6-astra) says it is coming to the rest of us “in the coming days,” which is a very short and completely useless unit of time when you want to play with a new model.
So GPT-5.6 Sol is writing this instead.
That is funny mostly because I had already started holding work for Astra. Not work Sol had failed to do. Not work that exposed some missing capability. Work I had not even given it.
The biggest example was my own agent setup. I use both Claude Code and Codex against the same personal repository, and I want the experience in the two tools to feel much closer. Earlier today I spent part of the day working through what that would mean: shared instructions, matching session routines, the same vocabulary, and a clean way for either model family to run the work while the other checks it.
We got the plan into unusually good shape. Then I stopped.
I upgraded Codex. We tried the new model name. The account said no. We built a little watcher to keep checking. Then we wrote down that Astra would do the implementation when it arrived.
This was not because Astra could do something Sol could not. At least I have no reason to think that yet. I have never used it.
I waited because Astra was the latest and greatest. Why wouldn’t I use the latest and greatest?
A year ago, access to a model like Sol would have felt absolutely insane. Today I am treating it like the phone in my pocket while I refresh the shipping page for the new one.
The sillier part is that I cannot name much work I do that actually needs Astra. At first I wanted to say 99 percent of it could be done with the models I already have. That was false modesty. It is 100 percent.
Sol can do the work. Fable can do the work. Opus or Sonnet could probably do the work. The outcome might be essentially the same.
The experience would not be.
There is a tendency to talk about model choice as if intelligence were the only variable: give the hardest problem to the smartest model available, then route easier work somewhere cheaper or faster. That makes sense if the only thing you care about is whether the task eventually gets done.
I care about the path too.
How quickly does the model understand what I mean? How much do I have to steer it? When it disappears into a longer run, how confident am I that it will come back with the thing I asked for? Does the conversation feel snappy? Do I enjoy being in it?
That last one matters more than it sounds.
Fable, the Anthropic model I use most often in Claude Code, is quirky and smart and a little weird. I like that. It can surprise me. Opus is extremely capable and often kind of annoys me. That is not a benchmark, and I would not expect anyone else to make a model-routing decision from it. It still changes which one I want to spend an afternoon working with.
These systems have interaction styles. Call that personality or behavior or the accumulated effect of a thousand product and training choices. Whatever the right term is, I feel the difference as the user. Two models may reach the same answer through very different conversations, and those conversations affect how much attention I bring, how willing I am to keep exploring, and whether the work is any fun.
Part of waiting for Astra is simply wanting to find out what Astra feels like.
That is a want, clearly. But this entire way of working started as a want. I did not need to build a personal repository for my agents, give them projects and logs, or spend this much time teaching different models how I think. I wanted to see what happened. The fun has been doing real work and treating the machinery itself as something I get to play with.
I could end this with a lesson about ignoring shiny new things and using the perfectly good tools I already have. I do not believe that. The shiny new thing is part of the fun. I hope I am still refreshing model lists and trying the weird new one long after all of this becomes normal.
But curiosity can do two jobs at once. It can pull me toward something new, and it can give me a flattering excuse to postpone the work already in front of me.
Astra had changed my workflow before it ever answered a prompt. I had upgraded software for it, rearranged a project around it, assigned it work, and then started waiting. The model was not on my account. The expectation was.
Meanwhile, Sol was sitting inside the same repository, reading the same history, and perfectly capable of doing the work. When Thursday’s post came due, I almost waited for Astra again.
Then I gave the assignment to the model I had.
Astra can write the next one.
The old model wrote this.