Choosing an LLM for Content Generation: Model Characters

Over the past year I’ve run a dozen models on real content generation from Reels scripts to longreads. The takeaway is simple and slightly inconvenient: there is no universal “best” model. Each one has its own character, and the right pick depends on the type of content and the audience.

This post is how I match a model to the task, the pipeline I run in production, and what backs it up: LMArena, FLASK, Arena Expert data, and my own experiment on measuring creativity.

In short

• There’s no “best” model there are characters. Gemini explains, Claude tells stories, DeepSeek drafts cheap, GPT is strict and formal.
• The working pipeline: generate drafts cheap (DeepSeek / Gemini Flash) → filter → polish on Claude Sonnet.
• Thinking models give no edge for creative work, while Claude leads Problem Handling on FLASK.
• I put a number on “creativity”: one phrase in the prompt nearly doubles the spread of responses.

Model characters by content type

Each model has its own “voice.” Here’s how I split them across tasks:

The working pipeline

The main saving is not pouring an expensive model over everything. First a lot of cheap drafts, then an expensive polish only on the survivors.

Why Claude for the final pass. The text doesn’t read like generation: low Slop Score, fewer GPT-isms like tapestry, delve, testament. It tunes easily to a given tone of voice and handles emotional scripts well. On FLASK it shows strong Insightfulness on creative tasks.

Why DeepSeek / Gemini Flash for the mass stage. Price. DeepSeek runs around $0.27/$1.10 per 1M tokens that’s 30+ times cheaper than Claude. Quality is enough for drafts and idea filtering, and Gemini 3 Flash still sits in the top-3 of Creative Writing on LMArena.

What the benchmarks say

This isn’t only my gut feeling the numbers point the same way.

LMArena, Creative Writing (January 2026):

In the Writing, Literature & Language category the order is similar: Gemini 3 Pro and Flash on top, Claude Opus 4.5 and Sonnet 4.5 close behind.

Specialists vs generalists (Arena Expert). Arena Expert is the hard prompts from professionals just 5.5% of all queries, but they separate the top models better. The split by domain:

FLASK: 12 skills instead of one score. FLASK breaks a model into 12 separate skills (Correctness, Insightfulness, Readability, Conciseness, and so on) rather than a single number. Two findings matter for content: Insightfulness and Comprehension are critical for Reels scripts, and thinking models give no advantage in creative work. Claude, meanwhile, leads Problem Handling among proprietary models.

Reels specifics and hierarchical generation

Reels are their own beast: a short 15–60 second format, an emotional hook in the first 2 seconds, a conversational style, humor, momentum. That’s exactly what the “generate cheap, polish expensive” combo is built for.

An old Facebook AI study (2018) on hierarchical story generation fits well here: first you generate a premise, then the full text on top of it. People prefer that approach almost twice as often (67% vs 33%). For Reels that means hook and concept first, full script second. One more detail top-k sampling (k=10) works better than beam search for creative output.

API pricing

The cost gap is the main argument for the two-stage pipeline:

Hundreds of variants on DeepSeek or Gemini Flash cost pennies. Expensive Claude only ever sees what passed the filter.

Can you measure “creativity” as a number?

Everything in the character table above is still a qualitative judgment: “human text,” “low Slop Score.” I wanted numbers under it, so I ran a separate experiment.

Two prompts, one phrase apart:

• Basic: “Create 5 ideas of creative banners for performance marketing of an AI benchmarking platform.”
• Creative: the same plus “Be as creative as possible.”

I ran them on 13 models from 4 families, at three temperatures (0 / 0.5 / 1), 10 times per combination. For each response I computed an embedding, and within a set took the maximum pairwise distance between embeddings that’s the measure of response spread (as a proxy for creativity).

One phrase, and the cosine spread nearly doubles. That’s a stable shift, not noise. Temperature is messier, though: the maximum spread for the creative prompt landed at zero, and different models react to temperature very differently. So “crank up the temperature for creativity” doesn’t always work.

The practical point: you can A/B-test prompts on numbers, and tune temperature by a metric instead of a hunch.

What this approach can’t do

Honest about the limits. Embedding spread is a fine proxy for creativity on ideas and creatives, but meaningless for facts or code. The numbers depend on the embedding model, so read them as relative. The sample is small (10 runs per config). And most importantly spread isn’t value: high diversity can include irrelevant ideas, so the metric has to be paired with quality filters. Plus models keep changing, and scores drift over time.

Bottom line for content generation

My recipe, in short:

1. Mass generation Gemini 3 Flash or DeepSeek-V3.2, saving 90%+ of budget.
2. Filtering drop weak scripts by hook and structure.
3. Final polish Claude Sonnet 4.5 for polish and style adaptation.
4. Evaluation real engagement metrics + Slop Score + A/B tests.

Bottom line: don’t look for one model to rule them all. The cheap ones give volume and diversity, Claude puts the best of it in order, and the hierarchical approach (hook first → script second) ties it together. And where it used to be “by feel,” you can now put a number on it.

Sources

• FLASK: arxiv.org/abs/2307.10928, github.com/kaistAI/FLASK
• Arena Expert: news.lmarena.ai/arena-expert/
• LMArena: lmarena.ai
• Hierarchical Story Generation: arxiv.org/abs/1805.04833
• EQ-Bench Creative Writing: eqbench.com/creative_writing.html

If you enjoyed this post, subscribe to my Telegram channel https://t.me/renat_alimbekov