There Is No Setting

When a model bills twenty five times more output tokens than its competitor for the same one sentence answer, the first hypothesis is that something was left switched on. That hypothesis is testable in an hour, and on this build it is wrong in a way that costs money to discover late.

The suspect arm came from a July benchmark of DeepSeek V4 Flash against GPT-5.6 Luna on a game move emission task: 28,144 mean output tokens per call, against about 1,125 for the reference arm. Rerunning that exact route fourteen hours later, on a byte identical prompt with the same grader, produced 28,031 mean output tokens, a difference of 0.4%, at $0.007896 and $0.007865 per call respectively. 99.8% of every billed output token was reasoning rather than answer. The reference arm moved 1,125 to 1,190 tokens across the same interval, so the instrument itself had not drifted between the two runs.

Delivery is what changed. Legal moves went from 7 of 12 to 12 of 12 and truncations from two to none, at the same spend per call, so the cost per finished move fell 1.7 times purely because failures cleared. The model did not think less; it managed to arrive somewhere, and arriving is a different property from being cheap. The earlier run's cost per finished answer was partly a measure of its own failure rate.

The mechanism sits in how the build was shipped. Its model card states that the release includes no chat template, so each serving provider implements the prompt format itself, and the reasoning control is not a sampler parameter at all: it is a block of text prepended to the prompt, and the documented default prepends none of it. The next level up begins "Reasoning Effort: Absolute maximum with no shortcuts permitted", and the level above that is more emphatic still. There is no setting between the default and the maximum, and nothing below the default. The expensive runs were already at that floor.

Both available levers were tested directly on two positions, holding the prompt fixed. Asking for low effort moved one position by 6% and left the other unordered against its own baseline across three configurations, landing at 7,305, 10,823 and 13,827 tokens with no monotone response, a spread of roughly 30% that reads as run to run noise rather than control. Disabling reasoning entirely does zero the reasoning channel, and both test cells then returned illegal moves: one emitted a 253 token stub that failed to parse, and the other ran all 32,000 tokens of its budget into the ceiling with no reasoning channel at all, having inlined the whole chain of thought into the answer body and never reached the required JSON. The thinking does not go away when it is switched off. It relocates into the part of the response that has to parse.

The unfinished half of this is worth stating plainly. The claim that the volume belongs to the build rather than to one vendor's serving currently rests on one provider measured twice, because the second genuine endpoint hard failed eight of its twelve calls after ten to twenty minutes each. Its surviving four calls averaged fewer tokens, and that is survivorship rather than a provider difference. Method, the cell by cell table, the control probe and the limits are in the companion methodology and raw data.

A vendor that ships a reasoning model without a chat template has delegated the size of your bill to whichever provider answers the call, and providers can disagree by an order of magnitude on byte identical requests.