On most Anthropic agreements, committed spend you do not consume simply vanishes at the period end. It does not roll forward, and it is not refunded. Here is why, and how a buyer protects against the loss.
A committed spend agreement with Anthropic is a promise to pay for a certain amount of usage over a term, usually a year, in exchange for a better rate than pay as you go. It is a good deal when your usage matches the commit. It is a quiet loss when it does not, because on most agreements the spend you committed to but did not consume does not roll forward and is not refunded. It simply disappears when the period closes. Buyers are often surprised by this, and the surprise is expensive. This article explains why unused commitment vanishes, what it costs, and the terms a buyer can negotiate to stop the leak.
When you sign a committed spend deal, you are trading flexibility for price. You tell the vendor you will spend at least a set amount over the term, and in return the vendor gives you a discount off the standard token rates. The discount usually deepens as you move up through the commit bands, from the smaller bands toward 1M and beyond. The logic is simple: the vendor gets predictable revenue, and you get a lower unit cost. The catch is the floor. You owe the committed amount whether or not your usage reaches it.
The reason is structural, not malicious. A commitment is revenue the vendor has already booked against the term. If you spend less than you promised, the gap is not a credit you are holding, it is a shortfall against a floor you agreed to. Most standard agreements are written so that the commitment applies within the period and does not carry into the next one. When the period ends, the meter resets. Anything you did not use is gone, because the contract never treated it as a balance to be carried, only as a minimum to be met.
This is why the default position favors the seller. If you overcommit and underuse, the vendor keeps the difference. If you undercommit and overuse, you pay overage on top, often at a worse rate. Either way, the structure is asymmetric unless you negotiate it to be otherwise.
Consider a buyer who commits to a number based on an optimistic forecast, expecting a new product to scale fast. The product ramps slower than planned, and by the end of the year the company has used a fraction of what it committed. The unused portion does not transfer to next year and is not returned. The buyer has paid full price for usage that never happened, and the better rate they negotiated is irrelevant because they did not reach the volume that justified it. The overcommit turned a discount into a penalty.
The mirror image is just as common. A buyer who undercommits to be safe blows through the floor mid year and pays overage at a punitive rate on everything above it. The lesson in both directions is that the commit number is the single most consequential figure in the agreement, and getting it wrong in either direction costs real money.
Unused commitment does not have to disappear. Several terms, negotiated up front, change the math. The most important are rollover, carryover, a ramp, and a true reforecast right.
A rollover or carryover term lets unused commitment from one period count toward the next, within limits. Even a partial rollover, say a cap on how much can carry and for how long, materially reduces the cost of a missed forecast. Vendors resist full rollover because it weakens the predictability that justified the discount, but partial terms are negotiable, especially on larger deals. Ask for it explicitly, because it will not appear by default.
Rather than committing to a flat annual number, structure the commit as a ramp that starts low and grows as your usage actually scales. A ramp aligns the floor with the real adoption curve, so you are not paying for volume in month two that you will only reach in month ten. This is the single best defense against overcommitting on a new product, because it lets the commitment follow reality instead of leading it.
A mid term reforecast right lets you revisit the commit if your usage diverges sharply from plan. If a product is delayed or a use case is cut, the right to renegotiate the floor partway through the term keeps a forecasting miss from becoming a full year loss. This is harder to win, but it is worth raising, particularly when the commitment is large enough that a miss would be material.
The deepest protection is an honest forecast. Most overcommitment comes from accepting a seller framed projection that assumes everything scales on time. Build your own consumption model from real usage data, with conservative and aggressive scenarios, and commit to a number you are confident you will reach rather than one you hope to. It is almost always better to commit slightly low with a good overage rate than to commit high and watch the unused portion evaporate. Overage at the committed rate, negotiated in advance, means growth is simply more usage at your price rather than a penalty.
The commit you need is downstream of how efficiently you use the model. Before you sign a number, optimize the workload. Routing across Opus, Sonnet, and Haiku so each task runs on the cheapest model that handles it well, caching stable context at up to 90 percent off, and moving asynchronous jobs to batch at 50 percent off can cut aggregate spend 40 to 70 percent versus running everything on Opus. A workload optimized before the negotiation needs a smaller commit, which means less exposure to the unused commitment trap in the first place. Commit to the optimized number, not the wasteful one.
It is worth naming where overcommitment comes from, because the cause is almost always the same. A seller presents a forecast in which everything scales on schedule, the new product launches on time, adoption climbs steeply, and usage reaches the upper bands within the year. That forecast justifies a large commit and the deeper discount that comes with it, and a buyer who is excited about the technology is inclined to believe it. Then reality intervenes. The launch slips a quarter, a use case is descoped, a team is reassigned, and the usage that was supposed to fill the commit never arrives. The discount that looked attractive is irrelevant, because the volume that justified it did not happen, and the unused portion quietly disappears at the period end.
The lesson is not that forecasts are useless, it is that the forecast you commit against should be yours, built conservatively from your own data, not the seller's optimistic version. A commit you are confident you will reach, with the upside protected by a good overage rate, beats a stretch commit every time. The asymmetry is the whole point: undershoot the commit and you pay overage, which at the committed rate is simply more usage at your price, while overshoot the commit and the difference is gone for good.
The treatment of unused commitment is rarely spelled out in plain words, so you have to look for it. Search the agreement for how the commitment period is defined, whether the commit is measured monthly, quarterly, or annually, and what happens to any shortfall at the close of each period. Look for language that says the commitment does not carry forward, or that any unused amount is forfeited, or simply the absence of any rollover provision at all, which amounts to the same thing. If the contract is silent on carryover, assume the default applies and the unused portion vanishes. The quietest clauses are often the most expensive, and a buyer who reads only the rate and the term will miss the mechanic that decides whether a forecast miss costs a little or a lot.
Consider a company that commits to a year of spend expecting a new feature to drive most of the consumption. The feature ships two quarters late. By the end of the term, the company has used a little over half of what it committed. On a standard agreement, the unused portion does not roll into the next year and is not refunded. The company paid full freight for usage that never happened, and the better rate it negotiated bought it nothing, because it never reached the volume that rate was meant to reward. Now run the same year with a ramped commit and a carryover term. The floor was low in the early quarters when the feature was not yet live, the carryover let the modest early shortfall flow into the next period, and the loss shrank from half the commit to almost nothing. Same business, same delay, two completely different outcomes, decided entirely by terms that were negotiable at signing.
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.