Cost control
Track AI Spend Across OpenAI, Anthropic, Bedrock & OpenRouter
No provider holds the company total. Building it requires four explicit transformations: normalize units, align time, resolve ownership, and retain confidence on every record.
AI spend arrives as several partial truths: an API report, gateway telemetry, a seat export, and a cloud invoice. A company total becomes reliable only after those records share the same identity, currency, and evidence model.
A provider usage page can explain its own slice of spend. It cannot resolve the same model under another name, restore ownership lost on a marketplace invoice, or say whether another source is settled. Tracking OpenAI, Anthropic, Amazon Bedrock, OpenRouter, and the rest is therefore a reconciliation problem as much as an integration problem.
The company total does not exist at any one provider
Finance asks what the company spent on AI last month and who owned it. OpenAI knows its share. Anthropic knows another. Bedrock is folded into AWS, while coding-tool seats may sit on a card statement. Each source can be accurate within its boundary. The organization-wide answer must be assembled across all of them.
A multi-provider total is useful only when each component keeps its original source and owner.
Source quality sets the ceiling on attribution
Start by recording the strongest grain and evidence state each source can supply:
| Source | Grain you can get | Best confidence available |
|---|---|---|
| Anthropic / Cursor | Per key, workspace, and member | Provider reported |
| OpenAI | Partial breakdown; supplement via API | Provider reported |
| Amazon Bedrock | Coarse; attribution lost via marketplace | Pharos estimate |
| OpenRouter / gateways | Per-key tags on every routed call | Pharos estimate |
| The invoice (any provider) | Whatever the contract itemizes | Invoice matched |
Tag shared traffic before the provider flattens it
Shared API keys collapse many teams and models into one billing line. An AI gateway can stamp each request with team, project, customer, and model tags before that context is lost. Those tags preserve attribution across a large stack. The gateway’s cost remains Pharos estimate until an invoice settles it.
Four transformations make the records additive
Process each provider in the same order:
- Resolve model identity. The same model is spelled a dozen ways across providers and dates. Fold every spelling to one canonical identity or the per-model view fractures.
- Convert currency. A non-USD invoice and a USD estimate cannot be added until they share a unit.
- Label confidence. Carry each figure’s state, whether matched, reported or estimated, through the sum, so the total can say how much of itself is settled.
- Reconcile per provider, then add. Match each provider to its own invoice first, then sum the reconciled totals. Adding before reconciling bakes every provider’s gap into one figure you can no longer pull apart.
The total should remain reversible
A useful total can be decomposed by evidence state, provider, team, and model. Finance can close the invoice-matched share. Engineering can inspect the expensive path. A customer margin can retain the records that support it.
If a total cannot be decomposed into its sources, the team cannot explain why it changed.
Begin with the two or three providers carrying most of the spend. Put them on one schema and reconcile each against its invoice. Then add sources without changing the method. When evaluating software, use a reconciled month to test whether the product can identify the provider behind a change and open the supporting record. Use that test to compare the available tools.
Frequently asked questions
- How do I track AI spend across multiple providers?
- Connect each provider's cost or usage records. Tag shared traffic at a gateway, normalize model identity and currency, preserve each figure's source state, and reconcile every provider to its invoice before adding the totals. The result is one company-wide number that can still be traced to its inputs.
- Which AI providers expose the most granular cost data?
- Anthropic and Cursor provide detailed records by key, workspace, and member. OpenAI provides a partial breakdown that its APIs can supplement. Amazon Bedrock loses developer-level attribution when marketplace billing reduces it to a coarse line. A gateway such as OpenRouter can retain request-level tags before that detail disappears.
- How do you get one AI spend total across OpenAI, Anthropic and Bedrock?
- Resolve provider-specific model names to one canonical identity, convert currencies, and label every figure as invoice-matched, provider-reported, or estimated. Reconcile OpenAI, Anthropic, and Bedrock separately, then add their reconciled totals. This prevents the company total from changing simply because one invoice arrived later than another.
- Should I use an AI gateway for cost tracking?
- Use a gateway when shared keys would otherwise flatten many teams or projects into one provider line. The gateway can tag each request by team, project, customer, and model. Its cost figures remain estimates until reconciliation, but its tags preserve the ownership detail the invoice may omit.
- Why not just add up each provider's dashboard total?
- Provider dashboards differ in grain, timing, and evidence. Adding them directly can combine a settled invoice, a current provider report, and a catalog estimate. Reconcile and label each source first; add the resulting totals afterward.