What the console gives you
- Spend and usage over time,sliceable by model and workspace. Your first stop for “why was this week expensive”, at daily granularity.
- Token composition, including cache reads and cache writes. Worth actually reading: Anthropic bills cache writes at a premium (1.25x input) and reads at a deep discount (0.1x), so two apps with identical traffic can have very different bills purely on caching behavior.
- Credits, auto reload, plan, and rate limits.The “why did my calls start failing” page: an exhausted credit balance looks like an outage if you have never seen it before.
- Invoices, the numbers your accountant reconciles against.
How the billing model itself works, tiers, cache math, and the line items, is covered in Anthropic billing explained.
When console numbers surprise you: the reconciliation checklist
- Recompute one day by hand. Token counts times current rates per million (Opus 4.8 at $5 in and $25 out, Sonnet 4.6 at $3 and $15, Haiku 4.5 at $1 and $5). The arithmetic is in the Claude API cost calculator.
- Check the cache split. Heavy cache writes with few reads means you pay the write premium without harvesting the discount, common with prompts that change slightly every call.
- Check the model mix. One code path quietly on Opus can dominate a bill of mostly Haiku traffic.
- Hunt unattributed traffic. Retries, background jobs, staging pointed at production keys. If you keep per request logs on your side, the diff against the console names the culprit fast.
The structural gaps, and what closes them
Gap one: it is not real time. Console data trails your traffic, which is fine for finance and useless for stopping an agent loop that is burning tokens right now. Velocity detection has to live at request time, in your own path.
Gap two: your customers do not exist in it. Finest grain is workspace and key. Whether the $49 account that hammered you all month is profitable is not answerable from the console, or from the Admin API cost report, which shares the same grain. Per customer cost only exists if you tag each request with your user id when you make it, the setup described in Claude API cost per user and, at the usage mechanics level, in tracking Claude usage per user.
Division of labor that works: console and Admin API as the authoritative record of what you owe Anthropic, request time tracking as the operational truth of who and what caused it. Weckr is the second half in two lines, including per user caps the console could never enforce.
FAQ
Where do I see my Claude API spend in the Anthropic console?
Under your organization’s billing and usage sections in the console at console.anthropic.com. You get spend over time, usage graphs by model and workspace, your plan and rate limits, and invoice history. Prepaid credit balances and auto reload settings live there too.
Why does my console usage not match my expected costs?
Usual suspects: cache tokens billed at different rates than you assumed (cache writes cost more than reads on Anthropic), a busier model mix than you thought, batch versus standard tier differences, and background or retry traffic you forgot. Reconstruct a day of cost from token counts times current rates, and the gap usually names itself.
Can the Anthropic console show cost per customer of my app?
No. The console’s finest grains are workspace and API key, and all your end users typically share one production key. Anthropic never sees your user ids, so per customer cost has to be attributed in your own stack at request time. That is a structural limit, not a missing feature.
What is the difference between the console and the Admin API?
Same underlying data, different access. The console is the human interface with graphs and invoices. The Admin API exposes usage and cost reports programmatically behind an sk-ant-admin key, grouped by model, workspace, key, and service tier, for your own dashboards and reconciliation jobs.
How should I split workspaces and API keys for better cost visibility?
One workspace per environment at minimum (production, staging, development), so experiments never pollute production spend, and separate keys per service or major feature if you want the console to attribute at that level. That is as far as provider side attribution goes. Finer than that, per feature within a service or per user, requires tagging in your own request path.
Keep reading
The console tells you the bill. Not the business.
Whether that bill was profitable depends on which customers caused it against what they pay, and that view lives in your stack, not Anthropic’s. Weckr builds it from two lines around your Claude client. See the per user margin screen on the live demo, or start from the AI cost and margin guide.