Trying to estimate Pickaxe costs for a high-volume build — need help thinking through the math

[drtomhanson]

I’m building a Pickaxe called Ask Coach Kelly for baseball and softball players. Here’s the shape of it, and I’d love guidance on what this is realistically going to cost me to run.

How it works: A user enters their results from a series of physical performance tests. Coach Kelly responds with (1) where they rank compared to peers and elite athletes, and (2) the specific exercises and drills they should prioritize. Mostly a single-turn interaction — prompt in, coaching response out — though some users may follow up with clarifying questions.

Knowledge base: Roughly 25–30 documents, around 5 pages each. All text. This is where the ranking benchmarks and drill recommendations live.

Volume projections:

• Total addressable base: ~2,000 teams × 15 players = ~30,000 athletes

• Realistic paid conversion: ~10% = ~3,000 paid members

• Usage frequency: roughly 3 sessions per user per month

• So I’m budgeting for ~9,000 sessions/month at steady state, ramping up from a much smaller start with a few teams

• Each session is one substantial response, maybe a short follow-up exchange

Business model: Bundled into an existing $20/month membership, not sold standalone.

Model: Open to recommendations. The documents carry the substantive content, so I don’t think I need a top-tier model — I’d rather optimize for cost as long as quality stays solid.

What I’m trying to figure out:

1. At ~9,000 sessions/month with a 25–30 document knowledge base, what’s a realistic monthly cost range? A back-of-the-envelope estimate is fine — I just need to know if I’m looking at $50/month, $500/month, or $5,000/month.

2. Which model would you recommend for this use case to keep costs in check without sacrificing coaching quality?

3. Does the size of the knowledge base meaningfully drive per-session cost, or is it mostly the response length and model choice that matter?

4. Do you offer volume pricing, a usage estimator, or any case studies from other high-volume builds I could reference?

5. Anything I should design into the Pickaxe from the start to keep costs down at scale (response length caps, retrieval settings, etc.)?

Appreciate any guidance you can offer. I want to go into this with eyes open before we roll it out broadly.

[nathaniel]

Hi!

This is a great question, thank you so much for asking it. We have a cost estimator tool you can use to map this out for yourself. We also have our model cost comparison page here.

The truth is, it matters a lot what model you end up using. Our advice is usually to start building with the newest, coolest, most cutting edge models. Then once you get it behaving how you’d like, reduce the cost of the model until you hit a good tradeoff point.

1. At ~9,000 sessions/month with a 25–30 document knowledge base, what’s a realistic monthly cost range? A back-of-the-envelope estimate is fine — I just need to know if I’m looking at $50/month, $500/month, or $5,000/month.

Our model cost comparison page actually shows the average cost per interaction across all of our users. So it’s a more inclusive proxy for cost and it might really be able to help you here. Let’s start by estimating that each session is about 10 interactions. That’s a fair benchmark for your case, some may be longer but others will be shorter.

That’d put you ad ~90K interactions per month. Here’s the broad range.
Low: Gemini 2.0 Flash is about $0.028 per 100 interactions, so it’d come out to ~$25.20 all in across all usage for all users.
High: GPT-5.2 Pro is about $46.72 per 100 interactions, so it’d come out to ~$42,048 all in across all usage for all users.
Typical: In a typical case you might choose a model like GPT-5, which is about $1.10 per 100 interactions, so it’d come out to ~$990 per month in model costs.

What we increasingly see is, you’d just pick whatever model you want, then use our new credits based billing feature. With that, you’d basically allocate them, say, $10 worth of usage. And if they go over, they pay for it themselves!

2. Which model would you recommend for this use case to keep costs in check without sacrificing coaching quality?

The model is like a base. You build quality on top of it. I’d start with GPT-5.

3. Does the size of the knowledge base meaningfully drive per-session cost, or is it mostly the response length and model choice that matter?

The amount in the knowledge base doesn’t effect cost. The amount you actually draw from the KB each generation does, but MUCH less than model choice. In the builder, in the configure section, in the token lengths section, you’ll see a breakdown of estimated cost.

4. Do you offer volume pricing, a usage estimator, or any case studies from other high-volume builds I could reference?

We offer a lot of tools for cost estimating as I’ve gone over above. We also have a variety of case studies, but they don’t go into cost directly (you can understand why). For more hands on assistance, you could consider a paid consultation with our team.

5. Anything I should design into the Pickaxe from the start to keep costs down at scale (response length caps, retrieval settings, etc.)?

The token lengths section is your friend here. But my advice would be to start without your eye on cost, and once you get the behavior you want, start dialing cost in. In the builder you can also see a breakdown of the cost of every message in the message insights section.