Committed spend vs on demand: the Anthropic math.

Every enterprise buying the Claude API eventually faces the same fork. Stay on demand and pay the standard rate for what you use, or commit to a level of spend and take a discount in exchange. The account team will frame the commitment as the obvious saving. Sometimes it is. Sometimes it quietly costs more than staying flexible. The difference comes down to math that most buyers never actually run, so this is a buyer side walk through of how the two models compare and how to decide between them.

We negotiate with Anthropic and study nothing else. The point of this piece is to put a clear number in front of both procurement and engineering, so the decision rests on arithmetic rather than on whoever has the louder opinion in the room.

What each model actually is

On demand pricing is exactly what it sounds like. You are billed for the tokens you consume at the standard published rate, with no commitment and no floor. Your spend rises and falls with your usage. There is no penalty for using less and no reward for using more.

Committed spend reverses both sides of that. You promise a minimum spend over a term and Anthropic gives you a discounted rate in return. Your effective price per token drops, but you now carry an obligation. If you spend less than the commitment, you still owe the commitment. If you spend more, the overage is billed at whatever rate the contract specifies. The discount is real, and so is the obligation. The whole question is whether the first outweighs the second for your specific usage.

The breakeven is the heart of the math

The decision turns on a single comparison. Take your honest forecast of annual Claude spend. Multiply it by the committed, discounted rate to get the committed cost. Then take the same forecast and multiply it by the standard rate to get the on demand cost. If your forecast is reliable and sits comfortably above the commitment floor, the committed model wins, because every token costs less. If your forecast is uncertain or sits near the floor, the picture changes, because the commitment can force you to pay for tokens you never consume.

The trap is the confidence of the forecast. A buyer who is sure of high usage should commit. A buyer who is guessing should be cautious, because the commitment converts a guess into a contractual obligation. The discount only helps on the spend you genuinely reach. On the gap between your commitment and your real usage, the discount is worthless and the obligation is total.

Optimization moves the breakeven

Here is the part the standard framing leaves out. Your forecast is not fixed. Before you commit to anything, you can change the number itself through optimization. Routing traffic across Opus, Sonnet, and Haiku instead of running everything on the top model, caching repeated context at up to ninety percent off, and sending non urgent work through batch at half price together tend to cut aggregate spend by forty to seventy percent.

That changes the math completely. A buyer who commits to a tier based on unoptimized usage, then optimizes, ends up locked into a commitment far above their real consumption. The discipline is to optimize first, forecast the optimized number, and only then decide how much to commit. The commitment should sit against the spend you will actually have, not the inflated spend you have today.

The risk on each side

On demand carries one risk: price. You pay the full standard rate on every token, so a large, stable, predictable workload pays more than it needs to. The flexibility is genuine, but for a buyer with confident high usage it is flexibility they are paying a premium for and may not need.

Committed spend carries the opposite risk: obligation. If usage drops, if a project is cancelled, if optimization works better than expected, you can be left holding a commitment you cannot fill. Whether that risk bites depends entirely on the contract terms around shortfall, which is why those terms matter as much as the rate.

The terms that decide the real cost

Two clauses turn the commit model from a gamble into a controlled trade.

• Overage rate. If you exceed the commitment, is the overage billed at your committed rate or at standard pricing. Committed rate overage means growth stays cheap. Standard rate overage means success gets expensive.
• Unused commitment treatment. If you fall short, does the gap simply disappear, or can unused commitment roll forward into the next period as credit. A rollover converts the downside risk into a timing question rather than a pure loss.

A commitment with cheap overage and a rollover on shortfall is a very different instrument from one where overage reverts to list and unused spend burns. The headline discount can be identical while the real economics diverge sharply, so read these clauses before you read the percentage.

A simple way to decide

Pull it together into a sequence any buyer can follow. First, optimize, so your forecast reflects routed, cached, batched usage rather than today's raw spend. Second, build a forecast you would defend to your CFO, with a sensible low and high case. Third, set the commitment at or just below your low case, so even in a soft year you fill it. Fourth, negotiate overage at the committed rate and a rollover on any shortfall, so both edges are protected. Fifth, take the discount on the volume you are now confident you will use.

Done that way, the commit model captures the discount without converting uncertainty into obligation. The flexible on demand model remains the right answer only when usage is genuinely unpredictable or the workload is too new to forecast, in which case staying flexible for a term and committing once the pattern is clear is the disciplined move.

A worked comparison

Numbers make the trade concrete. Take a buyer with a stable workload they are confident will run all year. On demand, they pay the standard rate on every token, month after month, with no obligation but no discount. If they commit to that same volume, the discounted rate applies to every token, and across a full year the difference is the discount multiplied by the entire spend. For a large, predictable workload that is a substantial saving, and the obligation costs them nothing because they were always going to spend the money anyway.

Now change one fact. Suppose the same buyer is only seventy percent sure of that volume, because a major use case is still being validated. On demand, if the use case is cancelled, their spend simply falls and they pay only for what they used. Committed, if the use case is cancelled, they still owe the commitment, and the discount on the spend that did happen rarely covers the cost of the spend that did not. The arithmetic flips on the confidence of the forecast, not on the size of the discount. The discount is identical in both cases. What differs is the probability that you fill the commitment, and that probability is the real variable in the decision.

A staged approach when usage is uncertain

The choice is not always all or nothing, and the most disciplined buyers treat it as a sequence rather than a single decision. When a workload is new or volatile, start on demand. Pay the standard rate while you learn the real shape of the usage, optimize aggressively with routing, caching, and batch so the pattern you observe is the optimized one, and only commit once the data is solid. You give up some discount in the early months, but you avoid the far larger cost of committing to a number that turns out to be wrong.

Once the pattern is established, commit to the floor you are now confident you will exceed, and leave the variable portion on demand at the overage rate you negotiated. This blended structure captures the discount on the spend you are sure of while keeping flexibility on the spend you are not. It is almost always a better answer than either extreme, because it matches the commitment to the part of usage that is genuinely predictable and no further. The buyers who lose money are the ones who treat the fork as a single irreversible choice rather than a position they can adjust as the data improves.