Choosing the Right AI Model
ReplyLabs offers 140+ AI models from six providers. Here's a quick guide:
- GPT-4o-mini (OpenAI). Best all-rounder. Fast, affordable (~$0.27 per 1,000 rows), handles most tasks well. Start here.
- Claude Sonnet 4 (Anthropic). Best for nuanced writing. Produces more natural, human-sounding text. Slightly more expensive (~$8.13 per 1,000 rows) but worth it for customer-facing copy.
- Gemini 2.5 Flash (Google). Fast and cheap. Good for simple extraction and classification tasks.
- o3-mini / o4-mini (OpenAI). Reasoning models. Best for complex multi-step analysis, but significantly more expensive. Use only when simpler models produce poor results.
Writing Effective Prompts
The quality of your results depends entirely on your prompt. Proven techniques:
- Be specific about what you want.Bad: "Write an email." Good: "Write a 3-sentence cold email to a VP of Engineering. Reference their company's recent product launch. Ask about their CI/CD pipeline challenges. Sign off as Alex from DevTools Inc."
- Include the format you want.If you want a one-liner, say "Respond in a single sentence." If you want JSON, specify the structure.
- Give examples."Here is an example of good output: [example]. Generate something similar for the following prospect."
- Set constraints."Do not use exclamation marks. Do not mention competitors. Keep the response under 50 words."
- Reference columns with curly braces.{Company Name} will be replaced with the value from the Company Name column for each row.
Using System Prompts
The system prompt sets the AI's identity and rules for the entire batch. It's like giving the AI a job description before it starts work. Enter your system prompt in the "System Prompt" field (click to expand it if hidden). Example: "You are a senior SDR at a B2B SaaS company specialising in supply chain software. You write brief, specific, personalised cold emails. You never use filler phrases. You always reference something concrete about the prospect." Every row in your batch will follow these rules.
Understanding Costs and Tokens
AI pricing is based on tokens, roughly 4 characters per token, or about 750 words per 1,000 tokens. Each prompt has an input cost (your prompt + row data) and an output cost (the AI's response). The cost estimate shown in the sidebar before you run is calculated from the model's per-token pricing and the average size of your selected rows. Actual costs may vary slightly based on response length. You're never charged more than your available balance.