Anthropic integration
LiveClaude cost monitoring, by model, workspace, and token type
Claude's bill has a shape most cost tools flatten: five token types, two of them cache lines that can swing which model is cheapest. CloudQuell keeps that shape — spend by model, by workspace, and by token type, with a cache-hit rate per workspace — and puts it in the same ledger as your OpenAI, AWS, and Snowflake cost.
It connects with a read-only Console admin key, reads Anthropic's cost and usage reports daily, and backfills up to a year. This is the tool-side companion to our Anthropic cost-tracking article, which works the cache math in full.
What CloudQuell ingests
Dollars from the cost report — the billing source of truth — joined to token counts from the usage report, including the cache activity that decides a Claude bill.
- Cost, per model per day
- Daily spend by Claude model from the cost report, reconciled to your Anthropic invoice.
- Five token types
- Uncached input, cache writes, cache reads, and output — the four usage lines, plus the cost line — so a heavily-cached prefix reads as nearly free instead of full price.
- Per-workspace attribution
- Spend by Console workspace, archived workspaces included, matching however you've split product surfaces or environments.
- Cache-hit rate
- Computed from cache-read versus input tokens per workspace — the single most useful number for cutting a Claude bill, because it shows which prompts pay full freight.
- Service tier
- Batch versus standard, where the Batch API runs at roughly a 50% discount — visible as adoption headroom.
What the data looks like normalized
Token types survive the trip into the ledger, so cache activity — the thing that decides a Claude bill — is still legible.
| Normalized field | Anthropic source | Example |
|---|---|---|
| Date | cost report day | 2026-08-04 |
| Provider | constant | Anthropic |
| Model | usage report | claude-sonnet-5 |
| Scope | Workspace | support-assistant |
| Cost (USD) | cost report | $172.40 |
| Tokens | token types | 22M in · 3.1M cache-read · 480K out |
| Cache-hit | cache-read ÷ input | 78% |
How it connects
A read-only Console admin key, scoped by the endpoints CloudQuell calls.
- Credential
- A Claude Console admin key (sk-ant-admin01-…). Regular API keys (sk-ant-api…) won't work — only the admin key can read org-wide usage and cost.
- Access scope
- Read-only in practice: Anthropic admin keys can't be permission-scoped the way OpenAI's can, so CloudQuell only ever calls the read-only usage and cost report endpoints. The key is validated live on connect and stored in AWS Secrets Manager, not the database.
- Prerequisites
- An Anthropic organization admin to mint the key. This is for Claude Console (the API platform) — Claude Enterprise on claude.ai is a seat subscription, not an API bill, and isn't in scope.
- Time to first data
- The first ingest starts right away and backfills up to 12 months. Sync runs daily and re-pulls a trailing three-day window so late usage reconciles.
- What's out of scope
- AWS Bedrock (the AWS connector) and Google Vertex AI (the GCP connector) — this is the Anthropic API platform bill, not Claude accessed through a cloud.
What it unlocks
Claude spend on the same model, allocation, and alerting as the rest of your bill — with the cache detail intact.
One ledger
By-model, by-workspace, by-token-type Claude spend in the same dashboard, cost centers, and reports as your OpenAI, AWS, and Snowflake cost.
Cache-hit analysis
Cost-per-request and cache-hit rate per workspace, so the prompt paying full price when it could be cached is a number, not a hunch.
Anomaly alerts
Per-model baselines catch a runaway agent the afternoon it starts, routed to Slack, Teams, or email — a smoke detector in front of the provider's hard cap.
AI budget alerts →Allocation
Route Claude cost to the same teams and cost centers as everything else, so the AI line has an owner like every other dollar.
The cache-math companion
This page is about putting Claude on one ledger with a tool. If you're building the pipeline yourself, our Anthropic cost-tracking article works the cache-write premium and read discount in full — the math that flips which model is cheapest.
Anthropic API cost tracking →Common questions
- How does CloudQuell track Anthropic (Claude) costs?
- It connects with a read-only Console admin key and reads Anthropic's cost report for dollars and usage report for tokens, daily. Spend lands by model, workspace, and token type — with a cache-hit rate per workspace — in the same dashboard as your cloud, OpenAI, and Snowflake bill.
- Is the Anthropic admin key read-only?
- In effect, yes. Anthropic admin keys can't carry a read-only permission flag the way OpenAI's can, so read-only is enforced by CloudQuell only ever calling the usage and cost report endpoints — never anything that runs a model or changes your org. The key is validated on connect, stored in AWS Secrets Manager, and revocable any time.
- Can CloudQuell analyze my Claude cache-hit rate?
- Yes. Because it ingests cache-write and cache-read tokens as distinct types, it computes a cache-hit rate per workspace and a cost-per-request. That's the lever for a Claude bill: a stable prefix reused across many calls should read as nearly free, and the hit rate tells you which prompts aren't.
- Does this work with Claude on Bedrock or Vertex?
- Not through this connector — Claude accessed via AWS Bedrock shows up on the AWS integration, and via Google Vertex AI on the GCP integration, because the charge lands on the cloud bill. This connector is for the Anthropic API platform (Claude Console) bill directly.
Free under $10K/month of tracked spend. Read-only admin key, cache-hit analysis, flat monthly price.