GVGanivra

Build vs buy · AI cost tracking

A token counter takes a day. Cost per customer that stays right is a product.

Before you build AI cost tracking in-house, here is the work it takes to know what each customer costs against what they pay, and what it costs to keep that number right.

Build it ifusage can never leave your systems, metering is your core product, or you use one model on one provider with a flat price.
Buy it ifyou want margin per customer this month, use more than one provider or pricing tier, or sell plans, credits or usage.

What you would build

Seven parts, and none of them is the dashboard.

  1. 01

    A price catalog that stays current

    165 models from 9 providers, 191 dated price versions. Prices change on announced dates, and last month's calls must keep last month's price.

    3 price changes and 20 model shutdowns are scheduled now →
  2. 02

    Token rules for each provider

    Cache reads and 5-minute or 1-hour cache writes, Batch, Flex and Priority tiers on 69 models, long-prompt rates on 17, and thinking tokens billed as output. Each provider reports usage differently.

  3. 03

    Capture without double counting

    SDK wrappers, a REST API or OpenTelemetry, and event IDs that make retries and replays safe. Failed calls still cost money and need to show up.

  4. 04

    Attribution to customers and work

    Customer, plan, workflow and feature on every call, and multi-step agent runs with MCP tool calls joined into one execution.

  5. 05

    A join to what customers pay

    Subscriptions, credit packs counted at the price paid, top-ups, refunds and currencies, matched to the same customer IDs as the usage.

  6. 06

    Corrections and history

    Fixing a bad event without losing the original, repricing history when a rate was wrong, and an audit trail of who changed what.

  7. 07

    Reports people act on

    Margin by customer, plan and workflow, the cause when a customer goes negative, budgets and alerts, and the month-end AI spend close.

Your numbers

What would building cost you?

Move the sliders to your own estimates. The starting values are examples, not benchmarks.

Build it yourself · first year$51,231$34,615 to build, then $16,615 a year to keep accurate
Ganivra · first year$1,188Growth plan, $99 a month for up to 100,000 calls

Engineering time only, at your own rates. It leaves out hosting, on-call and what the team could have shipped instead.

Side by side

Five things to weigh.

Build itGanivra
Time to first answerWeeks, before the first margin reportSame day: one SDK call, a REST POST or OpenTelemetry settings
Keeping prices rightSomeone watches provider pricing pages and ships updatesCatalog maintained, with scheduled changes added before they apply
Accuracy edge casesFound one invoice at a timeCache, tier, long-prompt and thinking-token rules already modelled
Who owns itAn engineer, usually alongside their real roadmapGanivra, while your team ships product
Your dataStays in your systemsUsage metadata only: no prompts, no responses, no proxy in the request path

An honest answer

When building is the right call.

Ganivra is not the answer for every team. Build, or use what you already have, when:

  • You only need total spend. Your provider's dashboard and invoice already show it.
  • Usage data cannot leave your infrastructure. Ganivra receives usage metadata, never prompts, but it is a hosted service.
  • Metering is your product. If you sell billing or metering, own it end to end.
  • One model, one provider, one flat price. A provider usage export and a spreadsheet may be enough for now.

Questions

Build vs buy FAQ

How long does it take to build AI cost tracking in-house?

Multiplying tokens by a price takes a few days. Cost and margin per customer that stays correct is longer: a dated price catalog, provider-specific token rules, attribution, a join to revenue and reports. Put your own estimate into the calculator above; most of the cost is in keeping it accurate after launch.

Is there anything we should still build ourselves?

Yes: the customer and workflow IDs your app sends with each call. Only your application knows who a call was for. Ganivra prices the call and joins it to what that customer pays.

Does Ganivra replace our observability tool?

No. Tools such as Langfuse, Helicone and Datadog show what ran. Ganivra shows what each customer costs against what they pay. Many teams run both.

Does Ganivra sit in our request path?

No. There is no proxy. Your app calls OpenAI, Anthropic or Gemini directly and sends usage to Ganivra afterwards, so an outage at Ganivra never blocks a model call.

We want to build. Can we use your price data?

Yes. The same prices Ganivra uses are published as JSON and CSV in the open llm-prices repository on GitHub under CC BY 4.0, including scheduled changes and shutdown dates.

See your own margin per customer before you decide.

Connect one workflow in an afternoon with the Python SDK, the REST API or OpenTelemetry. 14-day free trial, no credit card, from $29 a month after.

Start free trialRead the integration guide