AI Inference Cost Modeler

A free tool for founders pricing an AI product. You enter your own numbers: the plan price, how many tasks an active user runs each month, the input and output tokens per task, and the model prices you pay per million tokens, with the date you checked them. It shows what one paid account costs to serve, what it earns, and what it keeps, both at normal usage and when usage spikes.

What it checks

• Inference cost per task and per active user, blended across up to three models by the routing share you set. A task billed per call, such as transcription, can take a direct cost per task instead of tokens.
• All-in gross margin per paid account: plan and overage revenue, minus inference, hosting, support and payment fees, minus the share of the free tier's cost each paid account carries.
• The same account under a usage spike you choose, two times normal unless you change it.
• The break-even price, the price that hits your target margin, and the usage level at which an account turns unprofitable.
• The overage price. If an extra task costs you more after fees than you charge for it, the tool says so.
• Price freshness. Any model price checked more than 90 days ago is flagged. Provider prices change, so re-check them.

The verdict is measured against your target

The plan is healthy if the all-in margin at normal usage meets your own target and an account still makes money at the spike. It is thin if it makes money but misses either test. It is underwater if it loses money at normal usage. The tool doesn't supply a target or an industry figure. You enter the margin your plan or your board assumes.

Why I built it around the heavy account

With a seat price, a SaaS product costs about the same to serve whether a customer logs in once or all day. An AI product doesn't: every task has a cost. The average account can look fine while the busiest ones lose money. So I want two answers before a price goes on the page. Does it meet the target at normal usage? Does an account still pay for itself when usage jumps? If the second answer is no, I'd fix the usage terms first, meaning included tasks, overage and caps, and leave the headline price for later.

Levers, ranked

The tool tests fixed what-if moves and ranks them by how much each adds to all-in margin per account: price up 10%, 20 points of task share moved to the cheapest model, 20% fewer tokens per task, 20% fewer included tasks, and 20% less free-tier usage. Moving work to a cheaper model only counts if that model does the tasks well enough, so test quality first.

Limits

• The tool has no built-in model prices. The example uses made-up prices labelled "example price, not a quote".
• The model is only as good as your token counts. Measured averages from your logs beat estimates.
• The spike and the usage break-even scale paid usage only. To model a free-tier surge, raise the free-tier inputs.
• It models your own inputs. It is not financial advice.

The same engine runs as an agent skill, ai-inference-pricing. For pricing an AI product more broadly, see Step 6 of the Marketing AI Products Playbook. For tiers and blended ARPU, use the Pricing & ARPU Modeler. For LTV, CAC and payback, use the Unit Economics Calculator.

Run this in your agent

The free AAJ agent skill AI Inference Cost & Usage Pricing does this job inside a coding agent, on your own data. Install it with npx skills add sarojkjha/aaj-marketing-skills --skill ai-inference-pricing.

Using Claude or ChatGPT instead? Add the AAJ engines as a custom connector, with no install: https://aaj-engines-mcp.sarojkjha10.workers.dev/mcp