Almost Nothing Is Worth 10x the Token Budget

§1The pitch, and why it is seductive

Token-maxing is the idea that you make an agent better by spending more on it: more tokens, more tool calls, a bigger model, more retries, more reasoning. Turn every dial up and quality follows. It is seductive because it needs no thought. You do not have to work out which tasks are hard, or build any routing, or measure anything. You just crank the budget and trust that money buys quality. We had a month of real per-action cost and outcome data sitting in our audit log, so instead of arguing about it we went and looked.

§2The work that costs the most is already done

Here is the fleet over 30 days, priced not by raw spend but by cost per thousand successful outcomes, which is the number that actually matters if you are asking whether more money buys more value.

The workhorse, Gemini 2.5 Flash, does 12,906 calls at a flat 100% success for eight cents per thousand. That is nearly half our entire spend, and it is already perfect and already almost free. Pour ten times the budget on it and you buy nothing, because there is no headroom left to convert. A job that already succeeds every time cannot be made to succeed more. So for the single biggest line on the bill, the answer to "is it worth 10x" is a flat no, and the data does not even blink.

§3The expensive things are expensive by rarity, not by starvation

Look at the other end. An image costs 225 dollars per thousand successes, nearly three thousand times the flash rate. Opus costs 81. Those look like the obvious candidates to either cut or feed. They are neither. That is four hero images and a single Opus call across a whole month. They are not a starved workload begging for budget; they are rare, high-value one-offs, already rationed correctly because we only reach for them when we mean it. "Ten-x-ing" them does not make each one better. It just makes ten times as many, which you want only if you actually need the output. Spend is not the lever on quality here. Need is.

§4So who is worth 10x

Not a model. Not a policy you can set once and leave. The only thing that reliably earns ten times the budget is the narrow set of decisions that are rare, high-stakes and hard to reverse, where a stronger model or a longer think genuinely changes the outcome. And that is a routing problem, not a budget one. You want to spot the hard one percent and send it to the expensive model, not hand the expensive model to the ninety-nine percent that a cheap one already nailed. Token-maxing fails precisely because it cannot tell the two apart. It pays the premium on the work that was already done, and the one decision that would have justified the spend gets averaged into the noise. The honest policy is the mirror image of maxing: default to the cheapest thing that works, and spend big only where you can point at the specific decision that earns it.

§5The wrinkle, and the real lesson

Our own numbers came with a trap in them, and it is worth showing rather than hiding. In the same month, Haiku logged a 27% success rate and our judge panel 87%, which reads like a quality collapse. It was not. Our Anthropic credit ran dry partway through the month, so those calls returned billing errors, and an error counts as a failed outcome. A value-for-money metric inherits your outages: a perfectly good cheap model with an unrelated billing problem looks worthless until you notice the failures are 400s, not bad answers. That is the real lesson under the token-maxing one. "Worth 10x the budget" is unanswerable by vibes and only half-answerable by a naive metric. You need the actual log, and then you need to read it honestly, because the log records everything that happened, including the part that was your own fault.