Track your API usage every day without a spreadsheet

The API usage of an AI account reads like a water meter. The number exists and your provider updates it constantly, but someone still has to go and read it. You can organize that reading on three levels, from the native dashboard to a homemade log. A two-minute routine is enough for most uses.

The point is not bookkeeping for its own sake. Usage you look at regularly reveals the template that keeps swelling, the automation that runs wild and the discount that never lands, weeks before the bill tells the story in its own brutal way. Measuring means seeing things coming. The rest of cost optimization rests on that quiet foundation.

01Key takeaways

• There are three tracking levels to choose from by volume. They are the provider dashboard, a dedicated observability tool and a homemade log.
• For a solo user or a small team, level 1 is almost always enough. The native usage page already breaks things down by day, by model and by project.
• Four metrics do most of the work. They are input tokens, output tokens, cost per call and the share served from the cache.
• The winning routine takes two minutes a week. You compare against a budget set in advance.
• Tracking measures while a cap protects. The two go together and neither replaces the other.

02Why you look before the bad surprise

AI bills that shock people have one thing in common, which is that nobody was looking. Drift is almost never sudden. A template that grows or a loop that talks too much leaves traces from day one. A statement checked every week catches them while they still cost cents. A surprise bill is a symptom of missing tracking, more than a risk built into pay-per-use billing.

Transparency is the central argument of the whole guide to AI without a subscription. At your provider, every token consumed is visible, dated and broken down, whereas an intermediate credit system only shows its own units. That visibility is worth something only if you use it. The three levels that follow are three ways of using it.

03Level 1, the provider dashboard

The OpenAI usage page

The platform console has a usage page that breaks consumption down by day, by model and by project, with an export option. The official help page describes the current screens as versions change. If you have sorted your uses into projects, reading it becomes easy. Each line tells the story of a website, a client or an experiment, so an anomaly jumps out without any calculation.

Other providers' consoles

The logic is the same everywhere. Anthropic breaks usage down by workspace and by key in its console. Prepaid providers display the remaining balance and how fast it melts, while multi-model gateways give details by model used and by calling application. Each time you open an account, the reflex is the same. Find the usage page, bookmark it and connect it in your head to the official prices side by side so you can turn tokens into money.

04Level 2, dedicated observability when volume justifies it

Beyond a certain amount of traffic, the native dashboard becomes short-sighted. It tells you how much, but it can't tell you why. Specialized observability tools trace each call with its template, its tokens, its cost and its latency. Langfuse leads on the open source side, with Helicone and LangSmith among the managed services. You can replay a request, compare two versions of a prompt and spot the pipeline that consumes without producing anything useful.

Langfuse is a sensible starting point because it is open source, self-hostable and free to start with. The real question remains your need. These tools are built for applications that chain programmed calls together, while a content creator working by hand is not their audience. If you don't know whether you need one, you don't need one yet. Level 1 will tell you when the time comes, through its own limits.

05Level 3, the homemade log in three lines of code

Between the two, a frugal path exists for tinkerers. Every API response carries the detail of its own consumption, with input tokens, output tokens and the cached share. Writing those fields to a file on each call gives you a custom log that works with any spreadsheet. It takes three lines of code and zero external services. The breakdown follows exactly the categories you care about.

Level	Tool	Who it suits
1. Native dashboard	provider console	solo users, small teams, manual use
2. Dedicated observability	Langfuse, Helicone, LangSmith	applications with high call volume
3. Homemade log	usage fields in responses	tinkerers, precise breakdown needs

06The metrics that matter

Four numbers are enough to steer. Input tokens come first, the item that swells silently as templates grow. Working on template weight is the answer when that curve climbs. Output tokens come next, more expensive per unit and sensitive to answers that ramble. Cost per call comes third, the first two divided by volume. It catches drift better than the monthly total does.

The share served from the cache comes last, the number that confirms or disproves your discount lever. A cached share that collapses signals an edited template or an unstable prefix. Checking the cache discount gives the diagnosis. Latency matters for interactive applications. For content production it remains a comfort and does not steer decisions. Understanding billed tokens gives you the keys to read all of these counters.

07Reading drift, three scenarios on the statement

The first scenario is the gentle slope. The input curve rises a few percent each week with no triggering event. That is the classic signature of a template building up in layers, an instruction added here and an example there. The remedy is a diet for the template rather than a hunt for a culprit. The gentle slope never triggers a threshold alert, which is exactly why the weekly routine is what catches it.

The second scenario is the one-day spike. One day comes in at double or triple its neighbors and then everything returns to normal. Cross-check the date with your activity, because a batch launched that day or a rush delivery for a client is nothing abnormal at heart. If nothing matches, look for the loop or the script that ran twice. The breakdown by project points to the missing piece in a single reading.

The third scenario is the vanishing discount. The total rises while volume stays flat, because the share served from the cache has collapsed. An edited template, a prefix that became unstable or a timestamp slipped in at the top are the known suspects. You can check them by rereading the latest changes made to your working templates. Three scenarios call for three readings. The statement almost always tells you why as well as how much, provided you read it with method.

08In a team, the breakdown comes first

In a team, tracking is decided when the accounts are created. Set up one key or one project per person, per client or per website, otherwise the global statement mixes everything together and explains nothing. That breakdown costs five minutes per member at the start. It makes every line of the statement attributable and every budget conversation factual. A team that skips this step condemns itself to global statements that nobody feels responsible for.

The routine is then shared. One person owns the monthly check-in and presents three numbers, namely the total against the budget, the line that moved the most and the proposed decision. That takes five minutes in a team meeting. A culture of numbers takes hold with no heavy process or extra tool. Decisions about models get made on data instead of personal preferences defended out loud.

09Estimating before sending, the other half of tracking

The statement looks back while the estimate looks ahead. Before launching a new automation or a big batch, ten minutes of calculation tell you what the statement will later confirm. You need the template weight from the token counter, the planned volume and the list price. Estimating a call before sending it explains the method. The gap between that forecast and the month's statement then becomes your best anomaly detector.

That forecast and statement loop is exactly the loop of a living budget. Comparing against a budget set in advance gives the statement its yardstick, without which the numbers float around with no verdict. A statement without a budget only tells you what you spent. A statement with a budget tells you where you stand. That is the only reading that triggers a concrete decision at the right time.

10The two-minute routine

A routine that lasts has to be short. Every week, spend two minutes. Open the usage page, scan the seven-day curve and check that no day exceeds the pace you expected for the week just gone. Every month, add five more minutes to compare the total with the budget, look at the breakdown by model and write one number in your logbook for the record.

This routine complements the safeguard and does not replace it. Tracking sees drift coming while the cap stops it from costing you. Setting the cap that protects you is a one-time job. The statement gets reread every week and steers the rest, with the tour of cost-saving levers showing what to adjust depending on what the numbers say.

11Three months of history, when tracking gains value

One month's statement informs you, while three months of history teach you something. The seasonality of your activity appears, so the busy month stops looking like a worrying anomaly. The real cost per piece of content or per client can be calculated on honest averages instead of a hand-picked month. The effect of each setting shows up with a date attached, such as caching switched on in March and the input curve dropping at the same point.

History also shows the underlying trend. With constant usage, price lists fall and your monthly notes prove it figure after figure, better than any article could. A column of twelve numbers in a notebook is enough. On the day you need to defend a budget, decide on a model change or bill a client for usage, that column of dated figures is worth its weight in gold and costs only one minute a month.

12The checklist for tracking that lasts

1. Bookmark the usage page of each provider you use. Make one discovery visit to locate the screens and the available filters.
2. Sort your uses into projects or separate keys, the free breakdown that makes everything else readable.
3. Set the appointment with two minutes a week on the curve and five minutes a month against the budget.
4. Write down one history figure each month. Three months of notes are worth every dashboard on the market.
5. Move to level 2 only when the native dashboard no longer answers the why behind a recurring drift.

On OpenTahr, where the account is free, BYOK keeps this reading where it already exists. It stays on your provider account, in tokens and in real money, broken down by your own split of keys. No in-house unit stands between you and your numbers. API usage is read at the company that collects the money, which guarantees you never depend on a middleman's thermometer.

13Frequently asked questions

How do I see my OpenAI usage?

In the platform console, the usage page shows consumption by day, by model and by project, with an export option. The official help page describes the current screens. Two minutes a week on that page are enough for individual use. The breakdown by project makes the reading immediate.

What are Langfuse and Helicone?

They are observability tools built for AI applications. Each call is traced with its prompt, its tokens, its cost and its latency, which lets you understand a drift instead of merely noticing it. Langfuse is open source and self-hostable. These tools target high-volume applications rather than manual use.

Which metrics should I track first?

Track four of them, namely input tokens, output tokens, cost per call and the share served from the cache. Input catches swelling templates, output catches rambling answers, cost per call catches drift per unit and the cached share shows whether the discount lands. Latency only steers interactive uses.

How often should I check my usage?

Spend two minutes a week on the seven-day curve, then five minutes a month against the budget. That pace catches drift early without turning into a chore you drop out of boredom by the third month. Email threshold alerts fill the gap between two visits. They warn you between check-ins, while the statement explains what they saw.

Transparency. The screens described match the consoles and official help pages consulted on June 13, 2026, cited as sources. Interfaces change. The official pages are the reference. None of the providers or tools mentioned pays us.

14Sources

1. OpenAI official help on checking your token usage
2. Anthropic official documentation on limits and usage
3. Langfuse official documentation (open source observability)