I Gave ChatGPT, Claude, and Gemini the Same Cold Email Brief. One Model Won by a Mile.
Cold emails live and die by one thing: whether the person on the other end keeps reading. I handed all three major AI tools — ChatGPT, Claude, and Gemini — the exact same brief and asked them to write a cold email from scratch. The differences weren't subtle. One model produced something I'd actually send. The other two produced something I've seen a thousand times already. By the end of this article, you'll know exactly which AI to use for cold outreach, why the other two fall short, and how to prompt any of them to punch above their weight class.
The Test: Same Brief, Three Very Different Emails
The brief I used was deliberately real-world. No fluff, no "write me a cold email" — that's how beginners prompt. Here's what I gave each model:
"Write a cold email from a freelance UX designer named Sara to the Head of Product at a B2B SaaS company called Flowdesk. Sara wants a contract project. She has 5 years of experience, has worked with two fintech startups, and her last project reduced churn by 18%. Keep it under 150 words. No subject line needed. Make it feel human, not like a template."
Same prompt. Pasted verbatim into ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro.
ChatGPT came back clean and competent. It hit the word count, mentioned the churn stat, and structured the email logically. But it opened with "Hi [Name], I hope this finds you well" — which is the cold email equivalent of showing up to a job interview in a Halloween costume. It technically works, but the signal it sends is all wrong.
Gemini went the other direction. It tried so hard to sound human that it got vague. The fintech experience was buried in the third sentence, the 18% churn stat was softened to "significantly reduced churn" — which removes the entire point of having a stat in the first place — and the CTA was a weak "I'd love to connect." Nobody books a call because someone "would love to connect."
Claude did something different. It opened with a specific observation about the recipient's world, used Sara's numbers without burying them, and ended with a concrete, low-friction ask. It was also the only one that stayed under 140 words without being prompted to trim.
Why Claude Won — and What It's Actually Doing Differently
Claude's edge here isn't magic. It comes down to one thing most people ignore: instruction-following fidelity.
When you give Claude a constraint — 150 words, feel human, no templates — it treats those as hard rules, not suggestions. ChatGPT treats them as guidelines. Gemini sometimes treats them as inspiration. This matters enormously in cold email writing, where every extra word is a reason for someone to stop reading.
Here's what Claude's version actually opened with: "Saw that Flowdesk just rolled out your new onboarding flow — the redesign is clean, but I noticed the drop-off on step three might be costing you activations." That's invented for the email, yes, but it's the structure that matters. It leads with their world, not Sara's credentials. That's a completely different psychological move than "I hope this finds you well."
This reflects something called reader-first framing — a principle great copywriters have used for decades. The AI that naturally applies it wins cold email almost automatically. Claude has been trained on an enormous amount of long-form writing and tends to internalize narrative logic in a way that shows up in outputs like this.
ChatGPT isn't bad at email — it's great if you give it more scaffolding. Try adding to your prompt: "Do not open with a greeting. Lead with a problem the recipient is likely experiencing." With that addition, ChatGPT's output gets much closer to Claude's. But the fact that Claude gets there without that extra instruction is a real, practical advantage when you're iterating fast.
The deeper lesson: the best AI for a task isn't always the most powerful AI — it's the one that respects your constraints and understands the goal behind the words.
How to Extract a 90% Send-Ready Cold Email From Any of These Tools
You don't always get to choose your AI. Maybe your company uses Gemini through Google Workspace. Maybe you're on a free plan. Here's how to get the best possible cold email from whichever tool you have access to.
Step 1: Never start with "write me a cold email." That prompt produces template-brain output every time. Instead, front-load context. Use this structure: "You are writing on behalf of [Name], a [role] with [specific credential]. The recipient is [title] at [company type]. The goal is [specific ask]. Constraints: [word count], [tone], [what to avoid]."
Step 2: Add a "do not" list. This is the single highest-leverage prompt upgrade most people skip. After your main brief, add: "Do not open with 'I hope this finds you well.' Do not use the phrase 'I'd love to connect.' Do not soften the statistics." You're not just asking for good output — you're closing the doors to bad output.
Step 3: Ask for three versions with different openers. Prompt: "Give me three versions of this email. Each version should open differently — one with a problem the recipient is facing, one with a specific compliment about their product, one with a bold question. Keep every version under 150 words." This takes 30 seconds and gives you A/B test material immediately.
Step 4: Run the output through a gut-check prompt. After you get the email, paste it back in and ask: "Read this cold email as a skeptical Head of Product who gets 50 emails a day. What would make you stop reading? What would make you reply? Be honest." Claude is particularly brutal with this — in the best possible way.
Step 5: Send it. Most people tweak forever. The best cold email is the one that actually leaves your drafts folder.
The Part Most People Get Wrong
Most people use AI to write a cold email and then send the AI's first draft. That's wrong, and it's why their reply rates are terrible.
The AI's first draft is a starting point, not a finished product. Even Claude's excellent version would need one more pass: personalization that no AI can fake. The specific thing you noticed about their product, the mutual connection you share, the podcast episode you heard them on last week — that's the 10% that makes the other 90% land. AI writes the structure. You add the soul.
The second mistake is treating all three tools as interchangeable. They're not. Claude is your best default for cold email. ChatGPT is better when you want to iterate fast and want multiple structural options. Gemini is strongest when you're pulling in real-time context — like recent company news — through Google search integration and need to weave it into the copy.
The third mistake is writing cold emails without a clear single ask. Every AI will default to a vague CTA unless you explicitly tell it what you want the recipient to do. "A 20-minute Zoom call this Thursday or Friday" converts better than "a quick chat." Tell the AI the exact ask, and it will write toward it.
Key Takeaways
- Claude 3.5 Sonnet: The best default AI for cold email — it follows constraints precisely and applies reader-first framing naturally.
- Instruction-following fidelity: The hidden metric that separates good AI output from great AI output — always test whether your tool is treating your rules as rules.
- Reader-first framing: Open every cold email with the recipient's world, not your credentials — the AI that does this automatically wins.
- The "do not" list: Adding a list of things to avoid in your prompt is the single fastest way to improve AI-generated email quality.
- Human personalization: AI writes the structure; you add the specific, real-world detail that no model can generate — that's the layer that gets replies.
What to Do Right Now
Open Claude at claude.ai, paste in this prompt: "Write a cold email from [your name], a [your role] with [your best credential]. The recipient is [specific title] at [type of company]. The ask is [one specific action]. Under 150 words. Do not open with a greeting. Lead with a problem they're likely facing. Do not use the phrase 'I'd love to connect.'" Fill in your details, hit send, and you'll have a better cold email in the next 10 minutes than most people write in an hour.