Short answer: no model is best at every part of writing a novel, and we don’t have a verified ranking to give you. Pick two or three models you can already use, give each the same passage and the same prompt, hide the names, and keep the draft that needs the fewest fixes. Test again when the task changes: a model that outlines well can still flatten your dialogue.
The table below shows what to compare for each task and a quick test to run. The rest of the guide covers the full blind test, a prompt you can reuse and a worksheet. There is no public run log behind our earlier twelve-model comparison, so this guide no longer ranks models. The prose examples are editorial illustrations, not captured model outputs.
Pick a Model by Task
| Your task | What to look for | A small test to run | A reason to reject the output |
|---|---|---|---|
| Outlining and plot logic | Connects choices, consequences and character knowledge | Plant one contradiction in a short outline; ask for the conflicting passages | Invents an off-page event to explain it away |
| First drafts | Moves the scene forward while respecting your voice and constraints | Continue the same passage with the same beats and word limit | Uses the word budget without changing the character’s situation |
| Dialogue | Distinct voices and subtext | Give two characters conflicting goals and a secret neither will admit | Explains both characters’ feelings instead of letting them conceal them |
| Polishing and revision | Makes useful edits without flattening deliberate choices | Ask for a revision that preserves two specific sentences | Changes the protected sentences or adds facts |
| Continuity checks | Finds evidence and distinguishes absence from contradiction | Ask who holds an object at three points in the supplied chapters | Answers without pointing to supporting text |
Choose candidates from the models available in your writing setup. NovelThread’s model selection guide explains the controls; the plans page lists the public model catalog. Catalog membership is not a claim that one model writes better than another. Record the exact version you use rather than relying on a family name.
To see sample drafts from specific models first, read our four-model fiction writing test. It includes the full prompts and outputs; the scores are editorial judgments across three scenes, useful as reference samples rather than a general ranking.
For smaller tasks, try the book title generator or character name generator. AI generation uses the free allowance shown in the tool; local name generation uses a browser word list.
Run a Repeatable Blind Test
Use a passage you wrote and know well. A useful comparison keeps the source text, task, constraints and length target the same. Start a fresh conversation for each candidate so earlier messages do not give one a different brief.
- Select a finished passage and write down the facts the continuation must preserve.
- Send the same prompt to each candidate. Keep the original prompt and output unchanged in your notes.
- Label the outputs A, B and C before reading. If you are testing alone, ask someone else to hide the names when practical; otherwise treat the result as an unblinded personal comparison.
- Repeat the task a few times per candidate. Record failures and retries as well as the best answer.
- Compare the outputs before revealing the model names. Choose for this task and manuscript, then test a different task separately.
Keep the response length comparable. Record any settings you can control, including temperature, output limit and reasoning mode. If a service does not expose a setting, mark it as unknown; do not assume matching defaults. This method helps you make a personal choice, rather than establishing a universal ranking.
Run the same blind test in NovelThread: select a passage to rewrite, open More options, set Drafts to 2 or 3 and tick Use different models. Each candidate comes from a different model; pick one before you click Reveal to see who wrote it. Every candidate counts as one generation, billed to your plan’s credits or your own provider. The steps are in Multi-Candidate Rewrites & Blind Pick, and the desktop app is on the download page for Apple Silicon Macs.
A Prompt You Can Reuse
Continue the scene below using only the facts and constraints provided.
TASK
Write roughly 350 words. Mara and Julian test each other’s loyalties
before the footsteps reach the latch. End with Julian putting out the lantern.
CONSTRAINTS
- Third-person limited, locked to Mara. Do not enter Julian’s thoughts.
- Show unease through action, concrete objects and pauses.
- Neither character explains the secret or their full motivation.
- Do not invent a spare key, a hidden exit or an off-page agreement.
- Output story prose only, without a heading or explanation.
ESTABLISHED FACTS
- Mara entered the harbormaster’s office to find the manifest.
- Julian is on watch. He owes Mara’s family a debt but risks his post by helping.
- The brass key is in Mara’s hand.
- Footsteps are approaching along the dock.
MY FINISHED PASSAGE
[Paste a short passage you wrote and have permission to send.]The scene and names are fictional test material. Before sending your own excerpt, check which service will receive it and use text you are comfortable sharing. Keep private test logs private unless you choose to publish them.
Compare What You Would Keep
Count failures before assigning a preference. An elegant passage that reveals the protected secret or breaks the viewpoint still needs repair.
- Constraint errors: list each changed fact, viewpoint break, forbidden addition or missed scene beat, with the exact passage.
- Usable prose: mark the sentences you would keep unchanged. Record retained words divided by total output words, using the same counting method for every candidate.
- Revision effort: time a complete editing pass and keep the revised passage. Edit counts alone can hide one very large rewrite.
- Dialogue and voice: note concrete examples of distinct voices, subtext, unwanted clichés and changes to your style.
- Time and cost: record the observed waiting time and actual charged usage for all attempts. Keep token counts, provider charges and NovelThread credits in their own units.
Download the fiction model test worksheet to record exact versions, prompts, raw outputs, blind labels and editing notes. It is a blank template, not a completed benchmark.
When publishing a comparison, include the run dates, model identifiers, available settings, selection method and representative outputs you have permission to share. Explain how you chose samples and what failed. Only make model-specific performance claims that those records support.
What Restraint Looks Like on the Page
These two passages illustrate editorial choices. They were written for this guide and are not evidence that one model outperformed another.
Use the Result in Your Next Chapter
Keep the model and prompt that worked for the task, alongside the passage you accepted. Recheck your choice when the model version changes or your needs change. A short dialogue test does not establish how well a model will handle an entire manuscript.
In NovelThread, use rewrite candidates and blind selection to compare alternatives, and keep confirmed character details in your story bible. If the comparison changed a motive, timeline or clue, update your notes after revising the manuscript.
Disclaimer: Written by the NovelThread team, which develops the product discussed here. Workflow examples are fictional and illustrative. Features, availability and prices may change; consult the linked documentation. AI output needs author review. Comparisons are editorial guidance, not independent benchmarks.