How to Cut Your LLM API Costs and Protect Your SaaS Margins

Key takeaways

• Agent loops, not chat, drive most spend: a single user task can fire 1,000+ calls and burn millions of tokens.
• Headline input prices mislead. Output tokens often cost 3x to 6x more and quietly dominate the bill.
• One founder who moved heavy backend work off a premium API reported cost per user run falling from about $0.42 to $0.05.
• The number that matters is blended cost per active user, not cost per call.
• Model a provider swap against your own usage before you touch production code.

Why are LLM API costs eating SaaS margins?

Building is cheap now. Running is not. A founder writing on Indie Hackers described launching a lead-enrichment micro-SaaS in a weekend, then watching the API bill consume about 70% of subscription revenue. The culprit was not chat. It was autonomous agent loops that read output, evaluate it, and run again, sometimes 50 times per task. At that volume a single heavy user can push monthly recurring revenue negative.

This is the trap of usage-based costs sitting under a flat subscription price: your revenue per user is fixed, but your cost per user is not.

The mistake: optimizing the input price

When founders compare models, they usually look at the input price per million tokens. That is the wrong anchor. In agent workflows the model reads a short instruction and writes long reasoning, tool calls, and revisions. Output tokens are where the money goes, and output is typically priced 3x to 6x higher than input.

So a model that looks cheap on input can still wreck your margin if it is verbose. The blended rate, input plus output weighted by how your product actually uses them, is the only number worth comparing.

How much can a model swap actually save?

The Indie Hackers writer moved heavy backend logic to a lower-cost model and kept a premium model only as an optional fallback. The reported result, which is illustrative rather than a guarantee for your workload:

• Cost per 1M agent tokens: from roughly $3.00 to $15.00 down to $0.14 to $0.50.
• Average cost per user run: from about $0.42 to $0.05.
• Gross margin: from around 30% to 85%.

The lesson is not 'use this specific model.' Prices and model quality move every month. The lesson is that the gap between providers is now wide enough to flip a business from unprofitable to healthy, and most teams never run the math.

How do you model the switch before you ship it?

Do it on paper, or in a simulator, before you change code:

1. 1Measure your real token mix per task: average input tokens and output tokens for a typical user action.
2. 2Multiply by each provider's input and output price to get a blended cost per task.
3. 3Multiply by tasks per active user per month to get cost per user.
4. 4Subtract from your price per user. That is your true gross margin per seat.

Run that for two or three candidate models. The cheapest headline price rarely wins once output-heavy agent loops are included. This is exactly the kind of side-by-side you can build in Calcaas to compare providers and stress-test margins before committing.

What about lock-in and reliability?

A single-provider stack is also a single point of failure. Routing heavy backend work to a cost-efficient model while keeping a premium model for the cases that need it gives you both better margins and a fallback if one provider has an outage or a price change. Model-agnostic architecture is becoming the default for cost-conscious builders, and the cost modeling above is what tells you which work belongs where.

Takeaway: Do not pick a model on its input price. Model your blended cost per active user, then choose: you can simulate the whole comparison in Calcaas before you write a line of code.