Which Claude model to use at every stage of your writing pipeline — and when paying for the premium model actually pays you back.
Four models, four jobs. The skill isn't knowing which model is "best" — it's knowing which stage of your work each one is built for.
The flagship — a new tier above Opus. Highest intelligence ceiling available.
Use for: The judgment calls that everything else depends on — picking the angle, designing the structure, drafting your highest-stakes pieces.
The workhorse. Clear, warm prose with few AI tics. Strong editorial judgment.
Use for: Drafting, structural editing, ideation — the default spine of every pipeline.
Fast, capable, economical. Excellent with tools and long source documents.
Use for: Research, fact-checking, copy editing — the volume work.
Fastest and cheapest. Built for simple, mechanical tasks.
Use for: Typo and grammar passes only. Never for prose you'll publish.
Price differences come from per-token rates × how many tokens the model uses. A smarter model at lower effort can cost the same as a cheaper model that overthinks. Stop asking "which model is cheapest" and start asking "what's the cheapest path to the quality bar this piece needs?"
Every model also takes an effort setting — how hard it thinks before writing. More effort = deeper reasoning = more tokens = more cost.
| Effort | When to use it |
|---|---|
| low | Mechanical tasks — copy edits, formatting, classification |
| medium | Bounded tasks — line edits, headlines, research summaries |
| high | The default for real writing work — drafting, ideation, judgment |
| xhigh | Long-arc structure — outlines and drafts of 3,000+ word pieces |
The single most important decision isn't which model — it's which tier the piece belongs to. Everything else follows from that.
Pieces where one output's quality outweighs throughput:
Everything else:
Routine ≠ low quality. It means the Opus 4.8 spine is already at the quality ceiling for this work.
A 1,200-word sales letter deserves the flagship treatment — every sentence is load-bearing for conversion. A 3,500-word internal explainer doesn't. Stakes decide the tier. Word count only decides how much effort to apply within the tier (3,000+ words → bump high to xhigh).
The premium model only earns its cost on judgment and generation stages — ideation, angle selection, outlining, drafting, structural editing. It never pays back on research, copy edits, or fact-checking.
Recurring formats favor consistency over ceiling. Your newsletter audience notices voice drift more than marginal brilliance. A stable model + locked voice brief beats a smarter model week-over-week.
Voice doesn't carry itself between sessions. Every stage is a fresh context. Bring a voice brief plus 2–3 anchor paragraphs from your published work into every drafting and editing session — or the prose drifts toward generic AI voice no matter which model you pay for.
Eleven stages from blank page to published. Toggle between tiers to see how the model assignments shift. Click any stage for the why.
The pipeline doesn't change models per platform — it changes shape. Pick a content type to see which stages to keep, which to skip, and where the piece lives or dies.
Everything above, compressed to what you need at the moment of choosing.
Fable 5 appears only on flagship pieces, only on judgment/generation stages (ideation, angle selection, outlining, drafting, structural edit, headlines). Never on research, line edits, copy edits, or fact-check.
The routine column is a complete pipeline on its own. Opus 4.8 spine + Sonnet for mechanical work = professional-grade output. If budget tightens, routine becomes the answer for everything.
Validate before committing the premium. These placements are reasoned, not benchmarked. Run one real flagship piece through both tiers, judge blind, and let the result decide.
Conversion copy is always flagship — landing pages and sales letters get the premium treatment regardless of length. Revenue per word, not words per piece.
Compress stages for social, don't downgrade the model for the part that matters. A LinkedIn post skips research and outlining — but the hook deserves real model attention, because the hook is the post.