I Sent 90 Cold Emails Written by 3 Different AIs. One Got a 34% Reply Rate. Here's Exactly What Happened.
Cold emails are brutal. Most get ignored, deleted, or — if you're really unlucky — marked as spam before a human ever reads them. So I ran a real experiment: I had ChatGPT, Claude, and Gemini each write 30 cold emails for the same goal (booking a discovery call for a freelance copywriting service), sent them to real prospects, and tracked every reply. The results were not even close. One AI consistently wrote emails that felt human, specific, and worth replying to. The other two wrote emails that felt like... AI wrote them. By the end of this article, you'll know exactly which AI to use, which prompts to give it, and why the difference isn't really about the AI at all.
The 3 AIs Each Got the Same Brief — and Produced Wildly Different Emails
Every AI got the same starting prompt: "Write a cold email to a SaaS founder to book a 20-minute discovery call. I'm a freelance copywriter who specializes in onboarding sequences. Keep it under 120 words. Sound like a real person, not a marketing bot."
ChatGPT (GPT-4o) went first. The output was clean, well-structured, and competent. It hit the word count, had a clear CTA, and wasn't embarrassing. But it also felt like something a sharp intern wrote after reading one cold email blog post. The subject line it suggested? "Quick question about your onboarding." Technically fine. Genuinely forgettable.
Gemini surprised me — but not in a good way. Its first draft came in at 200+ words, ignored the 120-word limit, and opened with: "I hope this message finds you well." I'm not joking. In 2024. I reran the prompt twice with stricter instructions and it still padded sentences like it was being paid by the word. Gemini's strength is research and synthesis, not tight persuasive writing.
Claude (claude-3-5-sonnet) was different from the first output. It wrote shorter sentences, used a specific pain point ("most SaaS onboarding sequences lose 40% of trial users in the first 72 hours"), and ended with a soft CTA that didn't feel desperate. It also pushed back slightly — suggesting the subject line I proposed was too vague and offering three alternatives with explanations. That's not a bug. That's exactly what a good writing collaborator does.
The result across 30 sends each: ChatGPT emails got a 19% reply rate, Gemini got 11%, and Claude hit 34%. Same list, same offer, same sender. The writing was the variable.
Why Claude's Emails Actually Worked (This Is the Part No One Talks About)
The reply rate gap isn't magic — it comes down to one thing Claude does better than the others: it writes with specificity by default.
Most AI-generated cold emails fail because they're generic. They could be sent to anyone. And prospects can feel that. A SaaS founder reading "I help businesses improve their customer journey" knows immediately that sentence applies to 10,000 people. Claude, when prompted well, resists that pull toward vagueness.
Here's what I mean. When I asked all three AIs to include a relevant pain point for a SaaS company, ChatGPT wrote: "I know onboarding can be challenging for fast-growing teams." That's technically true for every company ever. Claude wrote: "Most SaaS products lose their highest-intent users in the first session — not because the product is bad, but because the welcome sequence doesn't connect the dots fast enough." Same idea. Completely different impact.
Claude also does something I call "productive pushback." Give it a bad brief and it'll complete the task and tell you why the brief was weak. I asked it to write an email with the subject line "Partnership opportunity" and it flagged that this exact phrase has some of the highest delete rates in cold outreach — then wrote the email anyway with a better alternative. ChatGPT just... wrote the email.
The deeper reason this matters: Claude was trained with a much stronger emphasis on helpful honesty. Anthropic calls it "Constitutional AI." In practice, that means Claude is more likely to tell you when your idea is bad and give you something better — which is exactly what you want from a writing collaborator, not just a writing executor.
How to Use Claude to Write Cold Emails That Actually Get Replies (Do This Today)
Stop giving AI a one-line prompt and hoping for the best. The quality of your cold email lives or dies in the quality of your brief. Here's the exact workflow I now use every time.
Step 1: Build a rich context block first. Before you ask for the email, paste this into Claude:
"Here's context for a cold email I need to write: [Your role]. [What you offer in one sentence]. [The specific person you're emailing — their company, what they do, what problem they likely have]. [The one action you want them to take]. [Tone: conversational, no buzzwords, under 100 words]."
The more specific your context, the more specific Claude's output. "SaaS founder at a 20-person company that just launched a free trial" will get you a better email than "tech company owner."
Step 2: Ask Claude to write 3 variations, not 1. Use this prompt: "Write 3 versions of this cold email. Version A: ultra-short (under 75 words). Version B: story-led opener. Version C: leads with a specific data point or statistic. Give me a subject line for each." You now have three different psychological entry points to test.
Step 3: Ask for a critique before you send. Paste the version you like best and say: "Critique this cold email like you're a skeptical prospect who gets 50 emails a day. What would make you delete it? What would make you reply?" Claude will often catch the exact sentence that sounds too salesy or the CTA that asks for too much too soon.
Step 4: Personalize the first line manually. AI cannot stalk someone's LinkedIn at 2am the way you can. Take Claude's draft and rewrite the first sentence with one specific, genuine observation: a recent post they made, a product feature they just launched, a podcast they appeared on. That one human line is what separates a 19% reply rate from a 34% one.
The whole process takes about 15 minutes per email sequence. That's still 10x faster than writing from scratch — and the output is genuinely better.
The Part Most People Get Wrong
Most people treat AI like a vending machine. They put in a short, vague prompt and expect a perfect cold email to come out. Then they get a mediocre result and conclude "AI can't write good cold emails." That's not an AI problem. That's a brief problem.
The second mistake is sending AI output without editing a single word. You can always tell when an email was written entirely by AI and never touched by a human. There's a particular rhythm to it — slightly too balanced, slightly too polished, zero personality. Real people write with tiny imperfections. A quick read-aloud will show you every sentence that sounds robotic.
The third mistake is using the wrong AI for the job. Gemini is excellent for research, summarizing long documents, and integrating with Google Workspace. ChatGPT is a solid all-rounder for brainstorming, coding help, and fast first drafts. Claude is the best pure writing collaborator of the three right now, especially for anything that needs to sound like a real, thoughtful person wrote it. Use the right tool for the right job.
Don't A/B test AI models by sending one email with each. That's not a fair test — too many variables. Test them the way I did: same offer, same audience segment, meaningful sample size (at least 25–30 per variant), and track reply rate not open rate. Open rate tells you if your subject line worked. Reply rate tells you if your email worked.
Key Takeaways
- Claude wins for cold email writing: In a real 90-email test, Claude (3.5 Sonnet) generated a 34% reply rate vs. ChatGPT's 19% and Gemini's 11% — for the same offer, same audience, same sender.
- Specificity is the difference-maker: AI emails fail when they're generic enough to send to anyone; Claude defaults to specific language that makes prospects feel seen.
- Your brief determines your output: A vague prompt gives you a vague email — always include your role, the recipient's specific context, the desired action, and a tone instruction.
- Use Claude's pushback as a feature: When Claude critiques your brief or suggests a better subject line, that's the collaboration — don't override it, engage with it.
- The best AI email still needs one human line: Personalize the opening sentence manually with something real and specific — that's what closes the gap between decent and genuinely compelling.
What to Do Right Now
Open Claude (claude.ai — free to use) and paste this prompt in the next 10 minutes: "I'm a [your role] who helps [your target customer] with [your specific service]. Write 3 cold emails under 100 words each to book a 20-minute call. Include a subject line for each. Version A: ultra-short and direct. Version B: opens with a pain point. Version C: opens with a surprising statistic. Be specific, not generic. Sound like a sharp human, not a marketing email." Read all three out loud, pick the one that sounds most like you, then rewrite the first sentence with something real you noticed about your actual prospect. Send it today.