GPT-6.1 Sol seems really good. My first impression after using it is basically Astra at a fifth of the price. It still isn't competing on design taste with Claude 5.5 models, but it's killer when it comes to cost efficiency.
When GPT-6 Sol came out, it wasn't nearly as good as I thought it would be. It was still better than 5.6 Sol in my experience, cheaper and faster, and it had a place in my workflow.
But my favorite thing about it was that it let me use Astra more. I could run the cheaper model for a lot of tasks and save my usage for the better one.
Then OpenAI released 6.1 Sol on September 29, just a week after 6 Sol.
6 Sol was barely born and OpenAI has already killed it.
I gave it my DevDay photos and went to bed
After DevDay, I asked Sol to go through my Google Photos and put together a video. I wanted it to pick the good photos, spend more time on the people I met, and avoid filling the video with repetitive shots of keynote slides. There were videos mixed in too, so it needed to decide which of those were worth including.
I also asked for music that didn't need attribution, cuts to the beat, and a few versions with different amounts of material. I told it I was going to bed and wouldn't be answering questions, so it needed to make those decisions itself.
It put the video together from that one prompt.
That's the kind of thing I want to use an agent for. I had a bunch of photos and wanted a video worth sharing. I could explain what I cared about and let it do the work.
It also made a Minecraft mob in Blender and took on Hakm's horror animatic test at high reasoning. It seemed really good across the tests I was trying.
The price matters a lot
Here are the standard API rates per million tokens for prompts up to 272K input tokens:
| Model | Input | Output |
|---|---|---|
| GPT-6 Astra | $10 | $50 |
| GPT-6.1 Sol | $2 | $10 |
Those are OpenAI's current rates. So the fifth-of-the-price part is literal for standard uncached input and output tokens. The Astra-level performance part is my early impression of using it.
That doesn't mean every task costs exactly one fifth as much. Models use different numbers of tokens, and retries count too. But getting this close in the work I'm giving it, at those rates, makes it a lot easier to use Sol for more things.
I've already set 6.1 Sol as my default in Codex. I also asked for my automations running on 6 Sol to move to 6.1, with the reasoning effort knocked down one tier unless they were already on low. I'll see how that holds up, but I'm comfortable enough with it to start moving work over.
Claude still has better taste
OpenAI models still have a tendency to add a bunch of text I don't want. My design instructions basically tell them that almost any text they add, I'm going to ask them to remove, so they should remove the unnecessary stuff ahead of time. If there's existing copy they can reuse, they shouldn't write something new just to fill the page.
Sol 6.1 hasn't changed my view that Claude 5.5 models have better design taste. If that's what you're optimizing for, it's still a reason to use Claude.
Theo, who I met at DevDay, found Sol less spiky than Astra. It might not be as smart, but it was significantly less dumb, making fewer unexpected mistakes. That's his report. I haven't measured that difference myself.
I want to give Sol more work. I need more stuff to do with it.