ChatGPT vs Claude vs Gemini: Who Writes the Best Cold Email?
I Gave All Three AI Giants the Same Cold Email Brief — Here's What Happened
Cold email is one of the highest-stakes writing tasks you can hand to an AI. Get it wrong and your message lands in trash. Get it right and you're booking meetings on autopilot. I gave ChatGPT, Claude, and Gemini the exact same brief — same target audience, same offer, same constraints — and tested each output against five real metrics: personalization, clarity, hook strength, call-to-action sharpness, and how human it actually sounded. One model didn't just win. It won by a margin that made the other two look like they were still in beta. By the end of this article, you'll know exactly which AI to use for cold outreach, which prompts unlock its best output, and why using the wrong tool here is costing you real replies.
The Brief I Used — and Why the Setup Matters More Than You Think
Before we get to the results, you need to see exactly what I sent each model. The brief was identical across all three, word for word.
The prompt: "Write a cold email from a freelance UX designer to the head of product at a 50-person B2B SaaS company. The goal is to get a 20-minute discovery call. The designer has worked with companies like Notion and Linear. Keep it under 120 words. No fluff. Sound human, not salesy."
That last instruction — "sound human, not salesy" — is where everything breaks down or clicks into place. It's the hardest thing to get right in cold email, and it's the clearest signal of which model actually understands tone versus which one just follows surface-level instructions.
I ran each model three times and used the best output from each. No cherry-picking the worst. This was a fair fight.
The five metrics I scored on were: hook strength (does the subject line and opener make you want to keep reading?), personalization depth (does it feel written for one person or broadcast to thousands?), clarity (can you read it in one pass and know exactly what's being offered?), CTA sharpness (is the ask specific and low-friction?), and human tone (would this pass if a friend sent it?). Each metric was scored out of 10. Max total: 50.
The Actual Results — Scores, Examples, and One Clear Winner
Claude scored 44 out of 50. ChatGPT scored 38. Gemini scored 31. That's not a close race.
Here's a condensed version of what Claude produced:
Subject: Quick question about your onboarding flow
Hi [Name],
I noticed [Company]'s onboarding is three steps longer than most of your competitors — which might be hurting activation rates.
I'm a UX designer who's shipped onboarding redesigns for Notion and Linear. Both saw measurable drops in time-to-value after we simplified the first session.
Would a 20-minute call make sense this week? Happy to share what we found.
— [Name]
That's 72 words. Specific, credible, focused on their problem — not the designer's portfolio. The hook references something observable about the company. The social proof is dropped in naturally, not bragged about. The CTA is low-pressure with a clear timeframe.
ChatGPT's version hit most of the same notes but ran 30 words longer and used the phrase "I'd love to connect" — a phrase that flags immediately as template copy. It also opened with "I came across your company" which is cold email filler at this point.
Gemini's version was the most polished-looking — great formatting, confident tone — but it read like a LinkedIn message, not an email. It used phrases like "I believe my expertise aligns well with your vision," which is the kind of sentence that makes anyone with a full inbox click delete.
The Hidden Reason Claude Wins — It Understands Subtext
Here's what most comparison articles miss: it's not just about the output quality. It's about how each model interprets ambiguous instructions.
When I said "sound human, not salesy," Claude treated that as a constraint that runs through every sentence. ChatGPT treated it as a tone note. Gemini treated it as a style preference it could partially ignore.
This difference in interpretation shows up most clearly in how each model handles social proof. Dropping Notion and Linear into cold email is powerful — but only if it feels incidental, not like you're name-dropping. Claude buried the proof mid-sentence. ChatGPT opened a sentence with it. Gemini bolded it. One of those approaches feels like a reference. Two feel like a pitch.
Claude also made an inference the other two didn't make: it invented a specific, plausible problem ("onboarding is three steps longer than competitors") rather than using a generic pain point like "improve user experience." That invention is risky — you'd need to verify it before sending — but it shows Claude is reasoning about what a real reader would find compelling, not just pattern-matching to cold email templates it was trained on.
The mental model here is this: great cold email solves a specific problem the reader didn't know you could see. Claude understood that. The others wrote solutions before establishing there was a problem.
This is the gap that separates good AI output from great AI output — and it's why your prompt framing has to do more than describe what you want. You have to tell the model whose shoes to stand in.
How to Use Claude for Cold Email Starting Today
You don't need a complex workflow. Here's what works in under 15 minutes.
Step 1: Build your context block. Before asking Claude to write anything, paste in: who you're emailing, their company URL or LinkedIn description, what you're offering, your social proof (2-3 clients max), and your word limit. More context = better output. Don't skip this.
Step 2: Use this starting prompt — "You are a cold email copywriter who prioritizes brevity and relevance over personality. Write a cold email from [your role] to [target's role] at [company type]. The goal is [specific ask]. Reference this about them: [observation about their company]. Keep it under [X] words. Do not use phrases like 'I'd love to connect' or 'reaching out because.' Make the first line about them, not me."
Step 3: Ask Claude to explain its choices. After you get the draft, type: "Why did you write the hook this way? What assumption are you making about the reader?" This forces the model to surface its reasoning — and you'll catch weak logic before it goes out to a real inbox.
Step 4: Run a "delete test." Ask Claude: "Which sentence in this email could be deleted without losing the core message?" If it identifies one, cut it. Cold emails almost always have one sentence that's doing nothing.
Step 5: Send version A to 5 people, then rebrief Claude with your reply rate and ask for version B. Claude is genuinely good at iterating when you give it performance data. ChatGPT and Gemini can do this too — but Claude's rewrites tend to change more meaningfully rather than just shuffling words.
The Part Most People Get Wrong
Most people use AI to write their cold email and then send it exactly as-is. That's wrong. Here's why: AI cold emails — from any model — are trained on patterns. The reader's inbox has also seen those patterns. You need to break at least one rule in every email you send.
The most common mistake is asking for "a good cold email" without specifying what makes your situation unusual. Claude is powerful, but it defaults to best practices. Best practices are invisible in a crowded inbox. You need to tell it what's weird, specific, or counterintuitive about your offer or your prospect.
The second mistake is treating the subject line as an afterthought. In every model I tested, the subject lines were the weakest part of the output — because subject lines require creative risk, and AI defaults to safe. Manually rewrite every subject line. A/B test two versions. The body of the email matters less than whether it gets opened.
Finally, stop prompting for "professional" tone. Professional is code for forgettable. The emails that get replies in 2025 sound like they were written by someone who actually looked at your company for 10 minutes and had a specific thought. Prompt for curious, direct, and slightly informal — and watch what changes.
Key Takeaways
- Claude wins for cold email: It outscored ChatGPT and Gemini on personalization, tone, and hook strength — often by significant margins in a direct comparison.
- Context is everything: The more specific your briefing prompt, the bigger the gap between Claude's output and what the other models produce.
- Subtext matters: Claude interprets ambiguous instructions (like "sound human") as constraints that run through the whole email — not just surface-level style notes.
- AI is a first draft, not a final send: Every AI cold email needs at least one manual edit — especially the subject line, which all three models get wrong by defaulting to safe.
- Iteration beats perfection: Give Claude your reply rate data after 10 sends and ask it to rewrite. That feedback loop is where AI cold email gets genuinely powerful.
What to Do Right Now
Open Claude.ai, paste in this prompt and fill in your details: "Write a cold email from [your role] to [target role] at [company type]. Goal: [specific ask]. Social proof: [2 clients]. Observation about them: [something specific you noticed]. Under 100 words. First line must be about them." Run it, then immediately ask Claude "Which word in this email sounds the most like a template?" Fix that word. Send the email today.