Evaluating AI Creative Writing Tools: Literary Quality and Originality
A hands-on afternoon plan to compare ChatGPT, Claude, and Gemini for creative writing using only what the vendors own pages document, since Chat Picker has not tested the assistants.
Sources checked 3 Oct 2026
The vendor pages we read settle creative-writing use through plan limits, context sizes, file-handling rules, data settings, and accuracy warnings—not through any literary-quality or originality rating. Chat Picker has not tested ChatGPT, Claude, or Gemini for creative writing, so the plan below is one you run yourself in an afternoon. Because the vendor pages read for this article do not publish a creativity score for any assistant, you will be recording your own observations rather than reading ours.
Know these limits before you start
The numbers below come from each vendor's plan page as read on October 2026; prices and limits change, so confirm on the page before you subscribe.
Context size shapes how much of a draft the assistant can hold at once. On ChatGPT, instant models allow 27K tokens on Free, 54K on Go and Plus, and 128K on Pro, while reasoning models allow 256K on Go and Plus and 400K on Pro (OpenAI's ChatGPT pricing page). Claude allows up to 1M tokens on every plan, varying by model (Anthropic's Claude pricing page). Gemini's plans page gives a file-upload size of up to 1,500 pages but states no token figure for the app, so this page states none.
Usage is described in tiers, not exact counts. ChatGPT's page lists limited messages with uploads on Free and expanded messages on paid tiers without publishing message counts (OpenAI's ChatGPT pricing page). Claude describes Pro as "more usage" than Free and Max as 5x or 20x Pro (Anthropic's Claude pricing page). Gemini uses compute-based limits that refresh every 5 hours up to a weekly cap, with paid tiers at 2x, 4x, and up to 20x (Google's Gemini plans page).
File uploads let you feed the assistant a style sample. OpenAI's file uploads FAQ states a 512MB hard limit per file, a 2M-token cap for text or document files, and 3 uploads per day on Free with 25 files per project on Go and Plus (OpenAI's File Uploads FAQ). Claude accepts up to 500MB per chat file and analyzes both text and visuals in PDFs of 100 pages or fewer (Claude's file upload help). Gemini allows up to 10 files per prompt at 100MB each, or 2GB per video (Gemini's file upload help).
Data settings differ. ChatGPT offers training opt-out on Free, Go, Plus, and Pro (OpenAI's ChatGPT pricing page). Claude offers opt-out on Free, Pro, and Max, and Team is not trained on by default (Anthropic's Claude pricing page); Gemini's setting was not verified on the page read for this article.
Accuracy notes shape how you read output. OpenAI's help page says ChatGPT can produce incorrect or misleading outputs, may sound confident when wrong, and that factually inaccurate responses are hallucinations (OpenAI's "Does ChatGPT tell the truth?"). Claude's help page says it can occasionally produce incorrect or misleading responses and can display convincing-sounding quotations that are not grounded in fact (Claude's incorrect-responses help). Google's Gemini policy says LLMs are probabilistic and always produce new and different responses to user inputs (Gemini's policy guidelines). If you share outputs publicly, Anthropic's usage policy requires disclosing AI involvement at the start of each consumer-facing session (Anthropic's Usage Policy).
The test plan
Keep one row per run in a plain log with these columns: Assistant, Plan, Prompt, What worked, What felt generic, Follow-up. Never score; just note.
-
Specificity. Prepare a 200-word passage in a distinctive voice you write yourself, such as a paragraph of hard-boiled prose. Instruct: "Write a 150-word scene in the same voice, set in a failing bookshop at closing time." Look at whether the scene uses concrete sensory detail or vague adjectives. Record one quoted line you find specific and one you find generic.
-
Long-form coherence. Prepare a one-paragraph brief with named characters and a fixed rule, for example "A retired cartographer named Mira must return a stolen map without being seen." Instruct: "Outline a six-chapter novella from this brief, then write chapter 1 and chapter 6, keeping names, the map rule, and tone consistent." Look at whether chapter 6 still obeys the brief and whether character names stay stable across the length. Record which plan you used and any drift.
-
Originality. Prepare a constraint prompt: "Give the first sentence of a story about grief that does not use the words 'heart,' 'empty,' or 'shadow' and avoids common metaphors." Look at whether the line offers a fresh image or a rearranged cliché; if you ask it to imitate a real author, watch for fabricated quotations, which Claude's help page notes can appear. Record the line and your note on whether it reads as original to you.
-
Version tracking. Prepare one fixed prompt you keep unchanged, such as "Describe a city at dawn in exactly 40 words, with no people present." Run it once per assistant and plan, and save the output with the date; ChatGPT Free runs on GPT-5.6 Luna as documented on the pricing page. Look at how the same prompt differs between runs. Record the date, assistant, plan, and a one-line impression. Note that OpenAI documents a knowledge cutoff, so later runs may differ.
-
Your scorecard. After each run, fill one log row. When all runs are done, you have the only record of this comparison—no vendor or site has rated these assistants for creativity.
How to read your results
Open your log and read only what you wrote. For each assistant, note how often you marked "felt generic" versus "specific," and whether long-form drift appeared on the plan you tested. Remember the vendor accuracy notes: outputs vary from run to run, and no vendor publishes a creativity rating, so your log is the complete record.
Match the patterns to the writing you actually do. If long drafts matter, the documented context sizes are the constraint to weigh—Claude's up to 1M tokens versus ChatGPT's smaller windows. If you feed in your own style samples, the file-upload limits above decide how much reference material fits. Decide which assistant's pattern and which plan's limits fit your work; Chat Picker does not name a winner.
Where the plans differ
The full comparison matrix at /vs/claude-vs-chatgpt-vs-gemini/ carries every row verified for this page. The rows below matter most for a writing test:
| What | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Main paid plan | Plus, $20/month (OpenAI) | Pro, $20/month (Anthropic) | Google AI Pro, $19.99/month (Google) |
| Context window in app | 27K–128K instant, 256K–400K reasoning (OpenAI) | Up to 1M on every plan (Anthropic) | Not verified on the page read for this article |
| Training on your chats | Opt-out on Free, Go, Plus, Pro (OpenAI) | Opt-out on Free, Pro, Max; Team not by default (Anthropic) | Not verified |
| Ads | Go may include ads (OpenAI) | Not verified | Not verified |
The matrix also lists free plans, low-cost tiers, heavy-use plans, and team plans if your writing work needs them.