# What an AI agent costs me per run, and why my first prices were wrong — retained record

> Retained reddit record from reddit:SaaS.

Canonical: https://highsignal.app/data/records/2a379ef5d77da5bd

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:SaaS)  published 2026-11-02 14:33 ingested 2026-11-03 02:39

[original source ↗](https://www.reddit.com/r/SaaS/comments/1xlys3z/what_an_ai_agent_costs_me_per_run_and_why_my/)

## retained text



```
I'm building Skyplanr, a mission planner for DJI drones. You describe the job (a real estate orbit, a mapping grid) and an AI agent running on Claude plans the waypoints from terrain, building heights, airspace and weather, then a validator checks the plan before you export it. Every plan costs me real money in tokens, so pricing has been the hardest part. No paying customers yet, so these are notes from the middle, not a success story.

**What one plan cost at first.** When I measured production runs at the start of October, the 18 that finished had a median of $3.73 a plan and the worst was $7.54. Most of that was not the prompt. The agent re-reads the whole conversation on every model request (about 29 per plan), and some tool results were huge, for example 40 to 90 thousand characters of building height data per call. Output was 38% of the bill, cache writes 33%, cache reads 28%.

**What brought it down.** Smaller tool results. Not repeating an 11k character JSON schema in two tool definitions. Letting the agent look up one helper function instead of reading a whole source file. Routing simpler jobs to Sonnet while mapping stays on Opus. And "ask for changes" now starts from the existing plan instead of from scratch, which costs about $0.33. Over ten days of production in October, real estate plans averaged $0.92 and mapping plans $2.07, with a worst case of $4.78.

**The first prices were wrong anyway.** I launched at €49 a month for 10 plans or €7 for a single plan. Two comments on Reddit settled it. A drone pilot pointed out that Litchi, an app a lot of pilots already use, costs about £20 once. A founder in a Feedback Friday thread did the math and showed my monthly plan never beat the free tier plus single plans or packs; only the yearly version did, and only at 10 to 12 jobs a month.

**Now.** 3 free plans a month, €3 per plan, or €24 a month for 10, also sold yearly, plus plan packs that never expire. My one rule was not to lose money on an average plan. After Stripe fees, a €3 plan keeps about €1.56 on a real estate job and €0.51 on a mapping job. There are holes: a yearly subscriber who only plans mapping jobs and uses all 120 plans loses me about €28 over the year, and a 25 plan pack spent only on mapping loses about a euro. I've left them for now and I'm watching real usage. If mapping heavy accounts show up, one option is counting a mapping plan as two.

What I'd do differently: measure cost per run for each job type before setting any price, and lead with pay per use when customers' volume swings a lot from month to month.

For those of you pricing AI features with a real per-run cost: do you charge per run, put a cap inside a subscription, or count heavy jobs as more than one?

I'm the developer, so weigh that accordingly. skyplanr.com
```



document fetched 2026-11-03 00:29 · canonical https://reddit.com/r/SaaS/comments/1xlys3z/what_an_ai_agent_costs_me_per_run_and_why_my
