Choosing an AI Model

AI Models on Coffeescribe

Coffeescribe uses AI models from Google, Anthropic, OpenAI, Perplexity, and others through OpenRouter. We organise model access around a lock-vs-unlock mental model: Free and Pro share a single curated default — the CoffeeScribe Model — so the experience stays predictable and the cost stays stable. Creator unlocks a model picker.

Model Access by Tier

See What is the CoffeeScribe Model? for the full reasoning behind the lock + the structure of the Creator picker.

The CoffeeScribe Model (Free + Pro)

Both Free and Pro generate with the CoffeeScribe Model — our curated default, maintained for the best balance of writing quality and cost. We update the underlying model from time to time as new releases land or providers retire models; the brand label stays the same so you don't have to track which specific engine is behind it.

You won't see a model picker on the Advanced settings panel as Free or Pro — it shows a locked CoffeeScribe Model badge instead. This is by design: predictable output, predictable cost, no decisions you don't need to make.

If you want to pick your own model, that's the Creator tier — see below.

Creator's model picker

Creator unlocks a structured picker:

1. CoffeeScribe Model (recommended) — pinned at the top, pre-selected as the default. If you don't change anything, you generate with the same curated model Pro uses.
2. Recommended models — about a dozen curated alternatives we've personally verified for writing quality and citation behaviour. Each shows Writing / Research / Cost dot ratings (1–5).
3. All other models (advanced — at your own risk) — a collapsed disclosure that opens up the full OpenRouter catalogue (350+ models). Smaller models — for example, those in the 8B-parameter range — may produce noticeably lower-quality output, and some may fail to generate at all. We don't review or guarantee these; the disclosure carries an explicit at-your-own-risk warning.

The picker is alternatives, not a switch. The CoffeeScribe Model stays the safe default; the other tiers are there for users who know what they're doing and want something specific.

How we rate models in the Creator picker

Every recommended model is scored on three things — Writing, Research, and Cost. Scores go from 1 (lowest) to 5 (highest), and they're relative to the other recommended models.

• Writing — How good is the prose? Does it write naturally, keep a consistent voice, and stay coherent across long chapters?
• Research — How well does it find and use facts? Is it good at pulling in current information from the web?
• Cost — How many tokens does it use per section? 1 = very cheap, 5 = premium pricing.

You can tap the info icon next to any model in the picker to see its full ratings, strengths, weaknesses, and citation behaviour.

Note: The model's Writing and Research scores directly influence the Quality Score shown on public scribes in the library.

Recommended Models in Creator's picker

Best for Creative Writing

• Claude Sonnet 4 — Writing 5/5, Cost 5/5. Top of the writing benchmarks. Nuanced prose, great voice consistency. Premium price — use when quality matters most.
• Claude 3.5 Sonnet — Writing 5/5, Cost 4/5. Previous-gen Claude, still an excellent writer. Slightly cheaper than Sonnet 4.
• GPT-4o — Writing 4/5, Cost 4/5. Strong all-rounder with engaging prose. Good structure and depth.

Best for Research-Backed Content

• Sonar Pro — Research 5/5, Cost 5/5. Industry-leading web research. Best for non-fiction, factual content, and topics that need current data. Writing is solid but factual rather than creative.
• Sonar — Research 4/5, Cost 2/5. Lighter version — good research at a much lower price.

Best Value

• GPT-4o Mini — Writing 3/5, Cost 1/5. Surprisingly capable for the price. Great for structured content and educational material.
• DeepSeek V3 — Writing 3/5, Cost 1/5. Good at structured and technical writing. Excellent value.
• Qwen3 235B — Writing 3/5, Cost 1/5. Top open-source model. Competitive quality at the lowest price.

Budget / Quick Drafts

• Gemini 2.0 Flash — Writing 2/5, Cost 1/5. Very fast and cheap. Good for simple, straightforward content.
• Llama 4 Maverick — Writing 2/5, Cost 1/5. Budget workhorse. Fine for blog posts and documentation.
• Grok 3 Mini — Writing 2/5, Cost 1/5. Fast and cheap. Better at STEM topics than creative writing.

Technical & Long-Form

• Gemini 2.5 Pro — Writing 3/5, Cost 4/5. Very large context window — handles the longest scribes without truncation. Best for very long scribes and technical content. Not the strongest creative writer.

Web Search

All models support web search through OpenRouter. When enabled, the AI searches the web for current information while writing. This is great for:

• Topics with recent developments
• Factual content that needs up-to-date data
• Research-backed non-fiction

Web search is available to all tiers. It's on by default — toggle it off if you prefer the AI to rely on its training data (and to save tokens).

Tips for Creator users choosing a model

• Starting out? Stick with the CoffeeScribe Model — it's pre-selected and tuned for the best quality/cost balance.
• Want the best writing? Claude Sonnet 4 or Claude 3.5 Sonnet.
• Need facts and research? Sonar Pro with web search enabled.
• On a budget? GPT-4o Mini, DeepSeek V3, or Qwen3 — all score 3/5 writing at the cheapest price.
• Very long scribes? Gemini 2.5 Pro handles the most context.
• Quick drafts to edit later? Gemini 2.0 Flash or Llama 4 Maverick — fast and cheap.
• Browsing the advanced disclosure? Read the at-your-own-risk warning. Smaller models (8B and below) often produce noticeably weaker prose; some models fail to generate at all.