Your Claude usage logs hold the commit number, but they do not hand it over. Here is the buyer side method for turning raw consumption data into a defensible committed spend figure before you sign with Anthropic.
You are close to signing, the commit band is on the table, and the question that decides whether the deal is good or bad is simple: what number do you commit to. The answer is sitting in your usage logs, but logs do not translate themselves into a commitment. They are a record of what happened, not a forecast of what to promise, and the work of turning one into the other is where most of the value is won or lost. Do it carefully and you commit to a number you can defend and consume. Do it loosely and you either prepay for tokens you will forfeit or undershoot and pay overage you never negotiated. This is the bottom of funnel mechanics: how to read your logs correctly and convert them into a commit figure that holds.
Raw logs are noisy, and the first job is to separate the signal from the noise. Pull a representative window, ideally several months, and strip out what should not influence a forward commitment: one off backfills, load tests, a launch spike that will not repeat, and any workload you are about to retire. What remains is your true recurring consumption, and that is the only part of the history that should feed the commit. Teams that skip this step commit against a baseline inflated by events that will never recur, and overpay accordingly. The opposite error is just as real: stripping out a workload that is actually growing because it looks like an anomaly. Clean the baseline deliberately, document what you removed and why, and you have a starting figure you can stand behind when the account team asks how you got there.
Logs usually record tokens, and a commitment is denominated in dollars, so the conversion runs through the price of each token, which is not a single number. Input and output are billed differently, with output costing several times more, so a workload heavy in long outputs costs far more per request than its token count alone suggests. Model matters too, since the same tokens cost very differently on Opus, Sonnet, and Haiku. To convert correctly, break consumption down by model and by input versus output, apply the right rate to each slice, and sum. A blended average across all of it will mislead you, because your mix is not uniform and the expensive slices carry disproportionate cost. Get the conversion right at this granularity and your spend figure reflects reality rather than an averaged approximation that could be off by a wide margin.
A commit covers a forward term, so a backward looking baseline is only the starting point. Two adjustments matter most. The first is growth: if your usage is climbing, the historical figure understates what you will consume, and you need to layer a defensible growth rate on top, tied to the business metric that actually drives volume rather than a flat percentage pulled from the air. The second is optimization, which pulls in the opposite direction. If you have not yet routed routine work off Opus, added prompt caching, or moved asynchronous jobs to batch, your logs reflect unoptimized usage that will fall once you do the work. Committing to the unoptimized figure means committing to waste you are about to remove. Net the two adjustments together, growth up and optimization down, and you arrive at a forward figure that reflects the workload you will actually run, not the one the logs happened to capture.
Once you have a forward figure, the instinct is to commit to it directly. Resist that. The commit should sit toward the lower bound of what you are confident you will consume, not at the midpoint of your expectation, and the reason is the asymmetry built into most Anthropic agreements. Unused commitment typically disappears at the end of each period, so every dollar you commit above actual consumption is forfeited outright. Overage, by contrast, is just paying for what you use, and if you have negotiated the overage rate well it costs little more than the committed rate. That asymmetry means undershooting the commit is cheap and overshooting is expensive, so you should size toward the figure you are confident you will exceed and let overage catch the rest. A commit set at your honest expectation has a fifty percent chance of leaving money on the table. A commit set toward the lower bound almost never does.
The final step is to make the number defensible, because a commit figure you cannot explain is one the other side can move. Keep the full build up: the cleaned baseline, what you removed, the model and input versus output breakdown, the rates applied, the growth assumption and its business driver, the optimization adjustment, and the decision to commit toward the lower bound. This is not paperwork for its own sake. When the account team proposes a higher band, a documented build up lets you show exactly why your number is what it is and what would have to be true for theirs to make sense, which turns the conversation onto evidence and away from pressure. It also protects you internally, since finance will approve a commit they can trace far more readily than one that appears as a bare figure. A well documented commit number is the difference between negotiating from a position you can defend line by line and negotiating from a guess.
Before you take the figure into the room, run one more test: reconcile it against your actual billing. The model should reproduce your recent invoices when you feed it your recent usage, and if it does not, the model is wrong and the commit built on it will be wrong too. This reconciliation catches the errors that quietly wreck forecasts: a rate applied at the wrong tier, a workload counted twice, an input and output split that does not match what you were actually billed for. A model that ties out to the invoice to within a small margin is one you can trust, and one you can defend, because when the account team questions your number you can show that it reproduces your real spend. A model that does not tie out is a liability you do not want to discover mid negotiation. Spend the hour to reconcile, because it is the cheapest insurance available against committing to a figure that was never right.
It can feel like overkill to break consumption down by model and by input versus output, layer growth and optimization adjustments, and reconcile against billing, all to produce a single commit number. It is not overkill. The commit is usually the largest single decision in the whole agreement, often worth more than every other term combined, and an error of even a few percent in the figure compounds across the term into real money. A loose conversion that overshoots by a modest margin forfeits that margin for the life of the commit, since unused commitment typically disappears. A loose conversion that undershoots badly exposes you to unplanned overage. The careful conversion costs a few hours and removes both risks, which is why it is the last analytical step before signing and the one most worth getting right. The discount you negotiate afterward matters, but it matters far less than committing to the correct number in the first place, because no rate concession recovers the value lost to a commit that was sized wrong.
Translating usage logs into a commit number is the last analytical step before you sign, and getting it right is worth more than any concession won at the table afterward. We do this conversion with clients, then negotiate the overage rate and the unused commitment treatment that make the lower bound commit safe. Get a quote to run yours with us, and read the pillar guide, the token optimization playbook, for the optimization adjustments that shape the forward figure.
Get a quote for a bounded engagement. Fixed fee or gainshare, no risk to you.
Get a QuoteWeekly intelligence on Anthropic pricing moves and the buyer side counters that work.