Cut the waste first, so spending is a choice
You can't reason about ROI while the bill is full of waste, because you don't know which spend is buying value and which is buying nothing. So clear the waste first: cache repeated calls, compress bloated prompts, right-size models for simple tasks, and cap runaway output. None of these hurt the product — they remove spend that was never producing value in the first place.
Once the waste is gone, the bill reflects real work, and you can make deliberate trade-offs. Now an expensive call is a decision, not an accident. This is also where measurement pays off: with a usage CSV in token·flow, you can see your costliest prompts and confirm they're the ones that actually matter to the product, rather than just the ones that happened to grow.
Match model power to the value of the call
ROI thinking changes how you route work. A high-stakes call — the one that closes a sale, drafts a legal summary, or resolves a support escalation — can justify your most capable model, because the value dwarfs the token cost. A low-stakes, high-volume call — tagging, routing, simple extraction — should run on the cheapest model that's accurate enough, because there the token cost dwarfs the marginal value.
The mistake at both ends is uniformity: running everything on the frontier model wastes money on grunt work, and running everything on the cheapest model quietly degrades the experiences that drive your revenue. The win is a tiered routing strategy where model power follows the value of the outcome. Cutting waste funds the expensive calls that are actually worth it — which is what maximizing ROI really means.
frequently asked
- Isn't the cheapest model always the best for ROI?
- No. ROI is value over cost, not cost alone. Downgrading a call that drives revenue or saves real staff hours can cost you far more in lost outcome than it saves in tokens. Use cheap models for low-stakes, high-volume work, and spend on capable models where the outcome justifies it.
- How do I decide which calls deserve an expensive model?
- Weigh the value of getting that call right against its token cost. High-stakes, low-volume calls — closing a sale, a legal summary, a support escalation — usually justify a capable model because the value dwarfs the cost. High-volume, low-stakes calls should run on the cheapest model that's accurate enough.
- Where does cutting waste fit into ROI?
- It comes first. Until you remove the waste — repeated calls, bloated prompts, oversized models, runaway output — you can't tell which spend is buying value. Clearing it makes the bill reflect real work, and frees budget to spend deliberately on the calls that earn it. token·flow surfaces the waste so you can act on it.
- Can I measure LLM ROI directly?
- You can get close by pairing cost data with outcome data: cost-per-successful-action, cost-per-resolved-ticket, cost-per-conversion. The cost side comes from your usage export (which token·flow breaks down by prompt and model); the value side comes from your product analytics. The ratio of the two is your ROI.
Last updated June 15, 2026