ChatGPT vs Claude vs Gemini: Who Writes Better Cold Emails?
We Tested All Three AIs on the Same Cold Email Brief — The Results Were Brutal
We gave ChatGPT, Claude, and Gemini the exact same cold email brief: pitch a B2B SaaS tool that automates invoice processing to a CFO at a mid-size manufacturing company. Same context. Same constraints. Same goal — get a reply. The results were so different it was almost hard to believe they're all "AI writing assistants." One wrote an email that felt like it came from a real human who'd done their homework. One wrote something generic enough to get deleted in under three seconds. And one made a mistake that would have gotten the sender blocked. If you're using AI to write cold emails right now — or you're about to — you need to see this before you hit send on anything.
We Ran the Same Prompt Through All Three — Here's Exactly What Happened
The brief we used was deliberately specific, because that's how real cold email work actually looks. The prompt was: "Write a cold email from a SaaS company called Clearflow to the CFO of a mid-size manufacturing company. The email should be under 150 words, reference a real pain point (late payments from customers), propose a 20-minute call, and feel human — not like a template."
ChatGPT (GPT-4o) went first. Its email opened with: "Hi [First Name], late payments are costing manufacturers like you more than just cash flow — they're costing you hours every week chasing invoices manually." Clean, punchy, specific. It hit the pain point in sentence one, didn't waste a word, and the CTA was a soft ask: "Would a 20-minute call this week make sense?" Not pushy. Not salesy. Genuinely good.
Claude (Claude 3.5 Sonnet) produced something more nuanced. It opened by acknowledging the CFO's likely skepticism — "I know your inbox is full of pitches, so I'll be quick." — before moving into the pain point. That one line of self-awareness changed the entire tone of the email. It felt like a person who understood the reader, not software spitting out a template.
Gemini (Gemini 1.5 Pro) was the problem. It opened with: "I hope this email finds you well." That's it. That's the line. In 2024, after a decade of people begging writers to stop using that opener, Gemini led with it. The rest of the email wasn't terrible — but the damage was already done. Nobody who gets 100 emails a day is reading past that sentence.
The gap between the best and worst output here wasn't small. It was the difference between an email that books meetings and one that trains spam filters.
The Hidden Reason Claude Writes Cold Emails Differently (And Why It Matters)
Most AI comparisons stop at "which one sounds better." That's the wrong question. The better question is: why does one AI understand persuasion and another one doesn't?
Claude has been trained with a heavy emphasis on Constitutional AI — a framework that makes it think carefully about how its output affects the reader. In cold email terms, that means Claude naturally considers the perspective of the recipient, not just the sender. It's not just filling in a template — it's modeling how a busy CFO actually reads their inbox.
This is why Claude added that line about acknowledging the CFO's skepticism. Nobody told it to do that. It inferred that a CFO getting a cold pitch would be defensive, and it pre-empted that. That's a persuasion technique called inoculation — you address the objection before it forms. It's used by the best human copywriters on the planet, and Claude did it automatically.
ChatGPT is excellent at structure and punch. It understands that cold emails need to be short, scannable, and specific. But it sometimes optimizes for sounding good rather than feeling real. The email it wrote was excellent — but if you read a hundred ChatGPT cold emails, you'd start to notice a pattern. There's a cadence to it. A rhythm. Savvy readers will clock it.
Gemini's problem isn't capability — it's defaults. When you give Gemini a vague or broad prompt, it falls back on the most statistically common patterns in its training data. And the most statistically common email opener in the world is, unfortunately, "I hope this email finds you well." The fix is simple: you have to be more aggressive with constraints in your prompt. More on that in the next section.
How to Use All Three AIs to Write Cold Emails That Actually Get Replies
The real answer isn't "use Claude and ignore the others." Each tool has a role in a smart cold email workflow — and using all three together takes about 15 minutes and produces something genuinely better than any one AI alone.
Step 1: Use Claude to write the first draft. Give it the full context — who the prospect is, what their likely pain point is, what you're offering, and how long the email should be. Prompt: "Write a cold email from [Company] to [Title] at [Industry] company. They struggle with [pain point]. Keep it under 120 words, feel like a real person wrote it, and end with a soft CTA for a 20-minute call." Claude's instinct for empathy and reader awareness makes it the best starting point.
Step 2: Run the draft through ChatGPT to tighten it. Paste Claude's output and say: "Make this colder, sharper, and more direct. Cut any word that doesn't earn its place. Keep it under 100 words." ChatGPT is ruthless about word economy. It will cut the filler that Claude sometimes leaves in. The result is usually leaner and more scannable.
Step 3: Use Gemini to generate subject line variations. Gemini is actually very good at producing volume — multiple angles on the same idea. Prompt: "Give me 10 cold email subject lines for this email. Mix curiosity, specificity, and directness. No clickbait." You'll get 10 options in seconds. Pick the best two, A/B test them, and you're done.
Step 4: Personalize one line manually. AI can't do real personalization — it doesn't know that your prospect just hired a new controller, or that their company missed their Q3 numbers. A single personalized line at the top of an AI-written email lifts reply rates more than any other variable. Keep it to one sentence: "Saw that [Company] just expanded into European markets — congrats." That's it.
The whole process takes under 15 minutes. And what you end up with is better than anything a human copywriter charging $50/email would produce.
The Part Most People Get Wrong
Most people test AI cold emails by reading them. That's the wrong test. You don't reply to emails you read — you reply to emails that make you feel something. The right test is to imagine you're a burnt-out CFO with 200 unread emails and ask: does this email make me want to reply, or does it make me want to delete it?
The most common mistake is treating AI as a writer instead of a first-draft machine. People take whatever ChatGPT or Claude produces and send it directly. No edit. No personalization. No subject line test. The AI got you 70% of the way there — but that last 30% is the difference between a 2% reply rate and a 12% reply rate.
The second biggest mistake is using a generic prompt and expecting a specific result. "Write me a cold email for my SaaS product" is a bad prompt. It's like walking into a tailor and saying "make me some clothes." The more specific your brief — prospect's title, company size, pain point, your offer, word count, tone — the better every AI performs. Gemini especially needs constraints to avoid its worst defaults.
Stop optimizing for emails that sound like AI wrote them, and start optimizing for emails that perform. Those are completely different goals.
Key Takeaways
- Claude 3.5 Sonnet: Best for first-draft cold emails — its reader-awareness and empathy produce the most human-feeling output by default.
- ChatGPT (GPT-4o): Best for editing and tightening — use it to cut filler and sharpen any draft that's running long or soft.
- Gemini 1.5 Pro: Best for subject line volume — give it your email and ask for 10 subject line variations to A/B test.
- Prompt specificity: The more context you give any AI (title, pain point, word count, tone), the better the output — vague prompts produce generic emails.
- The 70/30 rule: AI gets you 70% of the way there; the last 30% — one personalized line, a tested subject line, a human review — is what actually drives replies.
What to Do Right Now
Open Claude right now and paste this prompt: "Write a cold email from [your company] to a [prospect's title] at a [industry] company. Their biggest frustration is [specific pain point]. Keep it under 120 words, make it sound like a real person wrote it, and end with a soft ask for a 20-minute call." Fill in your real details, take the output, run it through ChatGPT to tighten it, and compare it to the last cold email you sent manually. The difference will be obvious — and a little uncomfortable.