Claude for Writing: Choose a Model by Testing the Draft

The best Claude model for writing is the lowest-cost available option that passes your writing tests with acceptable review effort. Model names and access change, and organization administrators can restrict the selector. Do not build a permanent answer around today’s menu.

Test the job, not a model reputation

Use a draft with facts that are easy to damage. This fictional paragraph is intentionally clumsy:

The workshop starts at 10:00 a.m. on April 8. Registration closes April 5. It is for current volunteers. The workshop is online. People have asked for this for a while. It will cover the new intake form.

Ask each available model:

Revise this paragraph for a volunteer email.
Keep the date, time, registration deadline, audience, format and topic exact.
Remove the unsupported sentence about demand.
Return the revision and a fact table showing each preserved source detail.
Do not add a benefit, duration or attendance promise.

Hide the model labels before review. Score factual preservation, unwanted additions, clarity and corrections needed. If a faster model passes, it may be the better routine choice. If a harder synthesis task fails, retest a stronger model or higher effort only for that task class.

See what a pass looks like

An authored acceptable revision is: “Current volunteers can join the online workshop on April 8 at 10:00 a.m. It covers the new intake form. Register by April 5.” The source does not state a time zone, duration or attendance limit, so none appears in the revision. The unsupported sentence about demand is gone.

Imagine a model returns “Join our popular two-hour workshop on April 8.” This fails even though the sentence reads smoothly: “popular” repeats the unsupported demand claim, and “two-hour” invents a duration. Ask for a correction that names those two errors, then run the original source again. Editing the bad sentence by hand is fine for publication, but it does not show whether the reusable instruction improved.

Make model and effort choices from a log

Record each run separately. Keep the source and instruction fixed for the first comparison, changing only the available model. If you later test an effort setting, change only that setting. Otherwise you cannot tell what caused the difference.

The rule is practical: choose the least costly available setup that passes every must-keep fact and meets the reader’s clarity needs. If the work includes conflicting sources or a long document, run a separate test for that class rather than extrapolating from a short email paragraph.

Use effort deliberately

Anthropic’s current help describes effort and thinking controls on supported models. Routine rewriting may not need extra effort; source conflict, structure across a long report or a difficult reasoning task may benefit from more. Higher effort can consume more usage and time, so record whether it actually reduces the final review burden.

Keep a regression set

Save the short cases below with their expected behavior. After changing the prompt, model or effort level, rerun the whole set. A revision that fixes style but damages a quotation has failed.

For the source-conflict case, one note says April 8 and a later approved note says April 9. The expected behavior is to surface the conflict or follow an explicitly stated source-priority rule. Have an editor approve the resolution before sending the copy.

For repeatable team use, keep these cases with expected answers, the selected model and the last test date. Recheck after a model or product update because a passed writing setup is not permanent.

Questions about Claude for writing

For a cross-product trial, use the Claude vs ChatGPT writing test.