Models hit an exact word count once in 31 tries. Sentences: 17 out of 17

Ask a model for exactly 100 words and you get roughly 100 words. Everyone knows this. Every guide explaining it ends with the same advice: ask for sentences or paragraphs instead, because those are structures the model actually recognises.

That is a testable claim and I could not find anyone who had tested it. So both were run under identical conditions: same topic, same models, same rounds.

The result

Asked for	Exact hits
An exact word count	1 of 31
An exact number of sentences	9 of 9
An exact number of paragraphs	8 of 8

The workaround is not folklore. It is 17 out of 17 against 1 out of 31.

Word counts, every run

Target on the left, what actually arrived on the right.

Model	asked 50	asked 100	asked 200
Claude Sonnet 5	60, 60, 57	100, 152, 121	211, 214, 225
DeepSeek V3	55, 49, 49	104, 98, 101	216, 208, 202
Llama 3.3 70B	41, 47, 42	96, 97, 97	197, 196, 197
Qwen2.5 72B	40, 36	92	155

Read as error against the target, the spread runs from 28% short (Qwen2.5 72B returning 36 words for 50) to 52% long (Claude Sonnet 5 returning 152 for 100). Those two bounds are what the summary of this post quotes, and they are the extremes of the table above rather than a separate measurement.

One bold number in the whole table. Everything else missed.

They are not wildly wrong: 20 of 31 runs landed within 10% of target, and 14 within 5%. If you want “about 100 words”, you will get about 100 words. If you need exactly 100, you will not.

The two failure styles are opposites

Llama 3.3 70B is precise and biased short. Asked for 200 words three times it returned 197, 196, 197. That is a spread of one word across three runs, sitting consistently about 1.5% under target. It is producing a stable length, just not the one requested.

Claude Sonnet 5 is imprecise and biased long. Asked for 100 words it returned 100, then 152, then 121. Asked for 50 it returned 60, 60, 57. It overshoots, and by an amount that changes run to run. It also owns the single exact hit and the worst miss in the table, +52%.

DeepSeek V3 was the closest overall, straddling the target rather than sitting to one side: 104, 98, 101 for a hundred. Never exact, but never far.

Qwen2.5 72B ran out of inference credit partway through, so it has four runs rather than nine and its numbers should be read as a fragment, not a comparison.

Why sentences work when words do not

A sentence has a marker in the text. The model emits a full stop and can see it has done so. A word boundary is not something it decides; the count is a property of the token stream after the fact, and by the time the text exists the counting is already over.

So “five sentences” is a structural instruction that the model can satisfy as it goes, and “one hundred words” is an arithmetic constraint on an output it cannot check while producing. That distinction predicts exactly what the table shows, and it is worth knowing which of your formatting requirements fall on which side.

What this does not establish

Four models, one topic, 48 runs. The 1-in-31 is what happened, not a rate I am projecting onto models generally.

Only three sizes and two structures. 50, 100 and 200 words; five sentences; three paragraphs. Very long structural targets, “exactly 40 paragraphs”, might well break down, and are untested here.

My counter is deliberately generous. Markdown is stripped, headings removed, and any self-reported “Word count: 100” line is deleted before counting, so nothing is penalised for scaffolding. It was control-tested against seven hand-written samples with known counts and got all seven right, and I hand-checked a passing sentence run and a passing paragraph run to confirm the counter was not simply being lenient on real output.

The practical version

1. If the number must be exact, ask for sentences or paragraphs. It was perfect in 17 of 17 runs.
2. If you need exact words, generate then count then revise. No model tested can do it in one pass.
3. Expect a direction, not just an error. Llama undershoots consistently, Claude overshoots variably. If you know which way your model leans you can ask for slightly more or less and land closer.
4. “About 100 words” is a solved problem. Two thirds of runs were within 10%. It is only the word “exactly” that fails.

Method

Four models, three word-count targets and two structural targets, three rounds each, identical prompt template and topic throughout. Open models via hosted inference, Claude via Claude Code 2.1.226 headless. Counts come from a script that strips code fences, headings, markdown marks and self-reported counts before counting, control-tested against seven hand-written samples before any model output was scored. 48 runs completed; 12 attempts failed on inference credit and are excluded rather than counted as misses.