Claude's usage dashboard shows what each API key spent, so give every tool its own key

The dashboard that shows your spend is Usage and Cost in the Console

If you searched for a Claude dashboard because you want to know what Claude is costing you, you want the Usage and Cost pages inside the Claude Console, the developer site at console.anthropic.com. Those pages show how many tokens you used, on which model, through which API key, and what that came to in dollars.

The chat app at claude.ai has its own settings screen. On a Pro or Max plan it shows how close you are to your plan's usage limits. That is a meter on a flat subscription. It is not a bill, and it never shows anything a tool spent through an API key. We explained why these are two separate places in our guide to the two Claude dashboards and which one controls your API key, so this page skips that and covers the part that guide doesn't: how to read the numbers once you're in the right place.

That matters more than it sounds. Many owners open the Console after the first month of using an AI-enabled tool, see a single line of spend, and can't tell whether it came from the inbox assistant, the quote drafter, or a test somebody forgot to turn off. The dashboard isn't hiding anything. It just can't separate things you never separated yourself.

What the Usage and Cost pages can tell you, and what they cannot

The Usage page counts tokens, the small chunks of text Claude reads and writes. You can filter or group it by date, model, API key and workspace. Input tokens (what you sent) and output tokens (what Claude wrote back) are counted separately because they are priced separately. Output usually costs several times more per token.

The Cost page turns that usage into dollars over the same breakdowns. This is where you answer "what did we spend last month" and "which model is the expensive one."

Neither page will ever tell you:

• Which customer a request was about. Anthropic sees text going through a key, not your client list.
• Which step of an automation triggered it. If one key powers five automations, they show up as one line.
• Whether the output was any good. A dashboard full of usage is not a dashboard full of useful work.

Your credit balance, invoices and payment method live under Billing, not Cost. API usage is normally paid from prepaid credits. The Cost page shows what you used, and the Billing page shows what you have left and what you were charged. If those two numbers don't match on a given day, that is usually timing, not an error.

Some organization accounts also have a separate analytics view for Claude Code, Anthropic's coding tool. A body shop or a photographer can ignore it unless someone on the team writes software.

Give every tool its own key before you look at a single chart

This is the step almost everybody skips, and it decides whether the dashboard is useful. Since spending is reported per key and per workspace, the only way to see what each tool costs is to give each tool its own key.

A practical setup for a small business:

1. Create a workspace per area of the business if you'll have more than a couple of tools, for example "Front office" and "Experiments." Workspaces can have their own spend limits, so a test can't eat the budget for the tool your clients depend on.
2. Create one API key per tool and name it after the tool and the date, such as "CRM inbox – Nov 2026." Never reuse one key across two products.
3. Set a spend limit on the workspace at a number that would make you look twice. You can raise it later. Finding out about a runaway automation from a limit is much better than finding out from your card statement.
4. Keep a one-line note of which key lives in which tool. When you rotate or revoke a key, you'll know what breaks.
5. Check the Cost page grouped by key once a week for the first month, then once a month.

Fynl runs its AI on your own Claude or ChatGPT key with no markup, so everything it does shows up in your Console at the provider's price. Give it a key of its own and you'll see exactly what the CRM's drafting and automations cost, separate from anything else you run. The same advice applies to any tool that asks for your key.

One caution: a key is a password that spends money. If a freelancer or developer needs access, give them their own key in its own workspace so you can revoke it without breaking everything else.

Turn the monthly total into a cost per job

A monthly dollar figure on its own tells you little. $11 could be cheap or wasteful depending on what it bought. The useful number is cost per job, and you can work it out with a count you probably already keep.

Here is a worked example. Assume a Sonnet-class model priced at $3 per million input tokens and $15 per million output tokens. That was Anthropic's published rate for Sonnet 4.5 at launch, so check the current pricing page for today's figure. Say a typical email draft sends Claude about 2,000 tokens (the customer's email, a bit of thread history and your instructions) and gets back about 500 tokens of reply.

• Input: 2,000 ÷ 1,000,000 × $3 = $0.006
• Output: 500 ÷ 1,000,000 × $15 = $0.0075
• Total: about $0.0135, or roughly 1.4 cents per draft

At 300 drafts a month, that is about $4. A clinic drafting 1,000 replies a month would be around $13.50 on the same assumptions.

Notice that output is the larger share even though it is a quarter of the length. That's why a tool that writes long, padded replies costs more than one that writes short ones. It is also why moving routine jobs to a smaller, cheaper model shows up so clearly on the Cost page grouped by model.

To get your real number, take the month's cost for one key from the Cost page and divide it by how many times that tool did its job: emails drafted, quotes written, enquiries sorted. Your CRM or your sent folder will have the count. If your cost per job is well above a few cents for short text work, something is sending far more text than it needs to, usually whole email threads or long documents on every request.

Three readings people get wrong

A spike is usually a machine, not a person

When a day's spend jumps to ten or twenty times normal, the cause is rarely that someone chatted too much. Usually an automation looped: a rule that replies to an email, which triggers itself again, or a sync that reprocessed every old message at once. Group the Cost page by key to find which tool it was, then check that tool's automation history for the same day. A workspace spend limit turns this from an expensive surprise into a short outage.

A zero doesn't always mean nothing happened

If your tool is clearly working but the dashboard shows nothing, check three things. First, the date range and any filters. Second, whether you're looking at the right workspace. Third, whether the key belongs to a different organization entirely. People who sign up twice, once with a personal email and once with a work email, often end up with two Console organizations and look at the empty one. Usage figures can also take a little while to appear, so a request from a minute ago may not be there yet.

Hitting a limit is not a billing error

Anthropic places API accounts in usage tiers that cap how much you can spend per month and how fast you can send requests. You move up as you buy more credits over time. If a tool suddenly reports errors and the dashboard shows you right at a ceiling, you've hit a tier or workspace limit, not a bug. Anthropic's rate limits page lists the current tiers, and the Console shows which one you're on. Running out of prepaid credits does the same thing, which is why auto-reload with a sensible cap is worth turning on for any tool clients rely on.

If you use a ChatGPT key, or want the numbers in a spreadsheet

OpenAI's equivalent is the usage page on the OpenAI platform. The same advice carries over. OpenAI groups keys into projects, and projects can have budgets, so one project per tool gives you the same per-tool view described above. If you run both providers, each dashboard only knows about its own keys, and you'll add the two totals yourself.

If you want Claude spending to land automatically in a spreadsheet or bookkeeping tool, Anthropic offers a Usage and Cost API that returns the same figures as the dashboard. It needs an Admin API key, and Anthropic's documentation notes the Admin API isn't available on individual accounts, only organizations. An admin key can manage your whole account, not just send requests, so treat it far more carefully than an ordinary key. Never paste it into a tool you don't fully trust.

For most businesses with a handful of tools, this is more machinery than the job needs. Ten minutes on the Cost page at the end of each month, grouped by key, plus one division to get cost per job, answers the question the API would answer. Reach for the API when an accountant or bookkeeper genuinely wants monthly AI costs split by department without you exporting anything.

The one exception is if you resell work to clients and want to bill AI usage through to them. Then separate workspaces per client, plus the Cost page or the API, is the only way to get numbers you can defend on an invoice.

Questions people also ask

Sources

• Anthropic – Usage and Cost API documentation
• Anthropic – Rate limits and usage tiers
• Anthropic – Claude pricing
• Claude Console
• OpenAI platform – Usage