The new breakdown is straightforward but creates a massive gap between reading and writing:
Input Tokens:$4.00 per million tokens** Output Tokens:**$20.00 per million tokens
When you look at it this way, the cost of generating a response is exactly five times higher than the cost of processing the prompt. For most of us doing standard
RAG(Retrieval-Augmented Generation), the input cost is the primary driver because we are stuffing massive context windows with document chunks. However, if you are building agents that perform iterative "Chain of Thought" reasoning—where the model essentially talks to itself to solve a problem—those output costs will eat your budget alive.
In a typical real-world scenario, if you have a prompt that is 10,000 tokens long and the model responds with 1,000 tokens of reasoning, your cost isn't just a flat fee. You are paying $0.04 for the input and $0.02 for the output. That sounds cheap until you realize that in a recursive loop where an agent calls a tool, observes the result, and then re-processes the entire history, the token count compounds exponentially.
If you're currently using a more expensive frontier model for everything, this might be a solid middle ground for specific tasks. But if you are already optimized, you might need to implement some aggressive prompt engineering to keep those output tokens lean. I've found that being overly verbose in system instructions often leads to "chatter"—the model explaining what it's doing instead of just doing it—which is essentially just burning money at a $20/million rate.
For anyone building a production-ready AI workflow, I'd recommend setting up a strict token usage monitor immediately. You don't want to wake up to a massive bill because a loop went rogue or a model got stuck in a repetitive generation cycle.
A simple way to mitigate this is to move the heavy lifting of summarization or data extraction to a cheaper model and only use Sol for the final, high-reasoning step. It's all about balancing that latency-to-cost ratio.
{
"model": "sol-api-latest",
"usage_estimate": {
"input_rate_per_1m": 4.00,
"output_rate_per_1m": 20.00,
"currency": "USD"
}
}
Next Picking a major for Physics-Informed Neural Networks (PINNs) is →
these AI tool field notes, with plenty of directly applicable cases.