Provider and model
Each event identifies the model that handled the call, so cost changes do not disappear inside one blended total.
AI cost attribution
Find the feature, model, and customer behind every reported AI cost.
Add the server SDK around the model calls your product already makes. It records provider-reported usage and the dimensions you choose, then prices each event against a versioned server catalog.
What is recorded
Enough to calculate cost. Not enough to read the conversation.
Input, output, and cached tokens
Cached usage stays separate because providers may price it differently from uncached input.
Status, latency, and retries
Failed or retried application calls remain visible when the provider returns usage, so reliability cost is not hidden.
Your own dimensions
Send opaque account, user, feature, and version identifiers that match the way your product is operated.
How it works
One event becomes a traceable cost.
- Observe the model response. The backend wrapper reads the usage returned by the provider.
- Attach product context. Your server supplies opaque account, user, feature, and version IDs.
- Price it centrally. The Grow or Die server applies the price version effective for that event.
- Keep uncertainty visible. Unknown models or missing usage stay unpriced rather than receiving a guessed cost.
Pricing boundaries
Token totals are not a bill until every pricing mode is known.
Grow or Die keeps standard input, output, cache reads, cache writes, service tiers, and supported tool fees separate. If an observed category has no trusted price, the call stays unpriced instead of becoming free.
Standard and batch pricing
The event records its service tier. Standard, batch, priority, flex, and custom execution are never assumed to share a rate; an unsupported tier blocks the cost result.
Cache reads and cache writes
Cached input, five-minute cache writes, and one-hour cache writes remain separate meters. A provider-specific price is required for every non-zero meter.
Tool fees stay visible
A supported web-search or other provider tool meter is added as its own line item. An unknown billable tool makes total cost unknown rather than hiding the fee inside token cost.
Unknown means unknown
Missing final usage, an unknown model, or a pricing category without evidence returns no cost. Calls and tokens remain visible, but profit calculations fail closed.
Worked boundary
Token cost + supported tool fee = observed call cost.
If a call reports 1,000 uncached input tokens, 2,000 cache-read tokens, 500 cache-write tokens, and one web-search invocation, Grow or Die prices each meter independently. If the cache-write or tool price is unavailable for that exact provider and mode, the result is unknown—not $0 and not a partial total presented as complete.
Private by design
Prompts, responses, and model API keys are not sent.
Grow or Die needs usage counts and your opaque business dimensions, not conversation content. Cost is calculated on the server so price changes do not require every customer to upgrade an SDK.
Set up my product