AI-generated metaphor showing an OpenAI logo, computation block and reviewed paper cards balanced on a mechanical scale.
AI-generated editorial illustration. Not an official brand image.

Why this caught my attention

A model launch makes me think about fairly ordinary jobs: pulling figures from a report, putting them into a consistent table, or preparing a first explanation of a quarterly change. Those are the tasks I’m interested in connecting into a process. If I want to run that process regularly, the cost of each run starts to matter.

How I would use it

OpenAI’s model reference lists standard rates of $2 per million input tokens and $10 per million output tokens, as checked on 30 September. The company positions Sol close to Astra on several kinds of work. I would treat that as a reason to try it on financial data, rather than a reason to assume it will handle that data well.

One idea I would like to explore is a small n8n workflow that takes a public report, passes selected material to an AI step and returns a structured table for review. I would keep straightforward calculations in formulas or code, and use the model where the information needs interpreting or organising. This is a proposed design, not an integration I have tested.

What would make it useful is consistency. If one run returns revenue in millions and the next returns it in thousands, I still have a problem to fix. The cheaper model becomes interesting when I can repeat the process without spending the saved money’s worth of time cleaning up the output.

I also do not want every new release to turn into a reason to rebuild everything. A small comparison on the same report would tell me more than switching models across an entire workflow straight away.

Where I want to take this

As a finance student, I want to get better at choosing which parts of a task to automate. That means understanding the data well enough to describe the output I need, spot an odd result and decide where a simple rule is enough.

I would start by comparing a few extracted figures with a manually checked table, then record the corrections and review time. If Sol makes that process cheaper and manageable, I would consider using it more often. If it does not, I still learn something useful about the task. For me, that is a more practical way to follow AI progress: see what it lets me build, and whether I would actually want to keep using it.

Sources