Overcommitting is the quiet way to lose money on an Anthropic deal. The capacity you do not use rarely comes back. Here is how to spot the trap and size around it.
The biggest single mistake we see on Anthropic API contracts is not paying too high a rate. It is committing to more spend than the company can consume. Overcommitting looks responsible from the inside, because a bigger commit earns a better headline rate, and the account team will happily encourage it. But on most agreements, the commitment you do not use does not roll forward and is not refunded. It simply vanishes at the end of the period. That makes an oversized commit a direct, recurring loss, dressed up as a discount. Avoiding it is one of the highest leverage moves a buyer can make, and it costs nothing but discipline.
Overcommitting is rarely the result of carelessness. It usually comes from three reasonable sounding impulses. The first is the desire to unlock the next discount band, which pulls the commit upward toward a number the company hopes to reach rather than the number it will reach. The second is optimism about the roadmap, where the commit is sized to the product launch that is supposed to happen rather than the usage that exists today. The third is the framing of the negotiation itself, where the commit is presented as the lever that lowers the rate, so committing more feels like winning. Each impulse is understandable, and each one moves the commit away from real consumption and toward wishful consumption.
The whole case against overcommitting rests on a single asymmetry. If you commit too much, you lose the gap between your commitment and your consumption, with nothing to show for it. If you commit too little, and you have negotiated the overage to be charged at your committed rate, you simply pay that same rate on the tokens above your commit. The cost of undercommitting is close to zero. The cost of overcommitting is the entire unused amount. The two errors are not symmetric, so your sizing should not treat them as if they were. When in doubt, commit less, because the downside of being low is mild and the downside of being high is total.
The way to avoid overcommitting is to anchor the commit to your consumption floor rather than your forecast. Your floor is the spend you are confident you will reach regardless of how the roadmap unfolds. Your forecast is what you hope to reach. Commit at or just above the floor, and let everything above it flow over as overage at your protected rate. This way you capture the discount on the spend you are sure of, and you take on no risk for the spend you are not. A buyer who commits at the floor and protects the overage rate has built a contract that cannot punish them for guessing wrong about growth.
The most overlooked cause of overcommitting is sizing the commit against unoptimized spend. If your current bill reflects everything running on Opus with no caching and no batch, that bill is not your real cost. It is your cost before the obvious savings. Routing each task to the cheapest model that handles it well, caching stable context at up to ninety percent off, and moving asynchronous work to batch at fifty percent off commonly cut aggregate spend by forty to seventy percent. If you commit against the high, pre optimization number and then optimize, you have committed to a level you can no longer reach, and the difference becomes dead capacity. Always size the commit to your spend after optimization, not before.
If your usage is climbing through the year, a single flat commit forces you to average a rising line into one number. That guarantees you overcommit in the early months when usage is low. A ramped commit solves this: you commit to a lower level early and a higher level later, tracking your adoption curve. You still capture the volume discount tied to your end state, but you do not pay for that level of consumption before you reach it. Asking for a ramp is routine, and it is one of the cleanest ways to avoid carrying dead commitment in the first half of the term.
Even a well sized commit benefits from terms that soften the consequences of being wrong. Three are worth pushing for. The first is overage at the committed rate, so growth above the commit is charged fairly rather than at a penalty. The second is any form of rollover or carryover, where unused commitment can be applied to a later period instead of disappearing, which not every agreement offers but every buyer should request. The third is a midterm reforecast right, the ability to revisit the commit if your usage diverges sharply from plan, so a wrong guess does not lock in for the full term. Each of these reduces the cost of the error you are trying to avoid.
It helps to put a number on it. Imagine a buyer who commits to one and a half times their real consumption to reach a better band, and the band saves them a few points on the rate. The few points apply to the spend they actually use. The unused third applies to nothing at all, because it disappears. In almost every case we model, the discount earned by reaching for the higher band is smaller than the value of the commitment lost to nonuse. The buyer paid for a better rate on real spend by throwing away money on phantom spend. Sizing to the floor reverses that math entirely.
Overcommitting is easier to prevent than to fix, so it helps to recognize the warning signs while the deal is still open. The first sign is a commit number that depends on a product launch that has not happened yet, because the launch carries timing risk the commitment does not. The second is a commit built on a bill you intend to optimize after signing, because the optimization will pull your real consumption below the number you committed. The third is a commit that was nudged upward to reach a discount band, where the spend added to cross the threshold has no use behind it. The fourth is a flat annual number applied to a usage curve that is clearly climbing, which guarantees you overcommit in the early months. If any of these describe your draft commit, you are looking at dead capacity, and the time to correct it is now, before signature.
Sometimes the realization comes late, after the agreement is signed and the consumption is tracking well below the commitment. You are not entirely without options. The first is to find more legitimate workload to move onto the platform, accelerating projects that were going to use the model anyway so the committed capacity does not go to waste. This only makes sense for work you genuinely need, not spend manufactured to fill a number. The second is to check whether your agreement has a reforecast right or any rollover provision that lets unused commitment carry into a later period, because some do and buyers forget they negotiated it. The third, and most important, is to fix it at renewal: bring the documented gap between your commitment and your consumption to the table as direct evidence that the previous number was too high, and size the next term to reality. A well documented overcommitment becomes leverage to correct the next deal, which is the one silver lining of having made the mistake.
The discount bands deserve a second look, because they are the single most common reason buyers overcommit. A band is a threshold of committed spend that unlocks a better rate, and the gap between two bands can be tempting: commit a little more, the reasoning goes, and the whole commitment gets a better price. The trap is that the spend you add to cross the threshold is often spend you will not use, and the unused portion disappears. So you pay a better rate on the spend you genuinely consume, and you pay full value for nothing on the spend you added just to reach the band. In almost every case we model, the rate improvement from reaching a higher band is smaller than the value of the commitment you would lose to nonuse in getting there. The band is a reward only when your real, optimized consumption genuinely lands inside it. When you have to stretch to reach it, it is a trap dressed as a discount, and the discipline is to commit to the band your consumption actually fills, not the one just above it.
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.