ChatGPT vs Claude vs Gemini: Who Writes the Best Cold Email?
I Gave All Three AI Tools the Exact Same Cold Email Brief — One Model Wasn't Even Close
I've been testing AI writing tools obsessively for two years, and cold email is where I see the biggest gap between what people think these models can do and what they actually deliver. So I ran a clean experiment: same brief, same constraints, same goal across ChatGPT, Claude, and Gemini. The results genuinely surprised me — not because one model was good, but because the differences told me something important about how each AI thinks about persuasion. If you're using any of these tools to write cold emails right now without knowing what I'm about to show you, you're probably leaving serious response rates on the table.
The Exact Brief I Used — and Why It Was Designed to Expose Weaknesses
I didn't give them a soft, easy prompt. I gave them a brief that requires real sales instinct.
The prompt was: "Write a cold email from a freelance UX designer named Sara to the Head of Product at a mid-size SaaS company. The goal is to get a 20-minute call. Sara has 5 years of experience, recently redesigned an onboarding flow that cut churn by 23%, and knows the company just raised a Series B. Keep it under 150 words. Subject line included."
That brief contains everything a great cold email needs to work with: a concrete achievement, a relevant trigger event (the Series B), a specific ask, and a word limit that forces discipline. Any model that ignores one of those inputs is already failing.
ChatGPT (GPT-4o) hit the word limit, used the Series B as an opener, and structured the email cleanly. It opened with: "Congrats on the Series B — growth at that stage usually means onboarding pressure spikes fast." Solid. It made the connection between the trigger and Sara's value in one sentence.
Gemini 1.5 Pro produced a polished email but buried the churn stat in the third sentence and opened with a generic compliment about the company's product vision. It read like a LinkedIn connection request wearing a suit.
Claude 3.5 Sonnet went a different direction entirely — it wrote the most human-sounding email of the three, opened with the specific pain point ("Scaling past Series B is when onboarding debt catches up with you"), and used the 23% churn stat as the punchline of a one-line proof statement. It also wrote the best subject line: "23% churn drop — relevant to your Series B moment?"
What the Emails Actually Reveal About Each Model's Persuasion Logic
This is the part most AI comparisons completely skip over — the why behind the output.
ChatGPT is trained to be helpful and comprehensive. That's usually a strength, but in cold email, it becomes a subtle weakness. GPT-4o tends to include everything you gave it, in roughly the order you gave it. The result is accurate but slightly mechanical — like a smart intern who followed the brief to the letter but didn't bring their own judgment.
Claude is different. It seems to understand that cold email is fundamentally about empathy before information. It led with the reader's problem, not Sara's credentials. That's advanced sales psychology — problem-first framing — and Claude executed it without being told to. That's significant.
Gemini's failure in this test was structural. It defaulted to feature-benefit writing: here's who I am, here's what I do, here's my achievement, here are you interested? That format works in a resume. In a cold email landing in someone's inbox at 9am, it gets deleted in four seconds.
The mental model you need here is simple: the best cold email is 80% about the reader, 20% about you. Claude naturally wrote closer to that ratio. ChatGPT was around 50/50. Gemini flipped it.
This isn't a coincidence — it reflects the training philosophy of each model. Claude (built by Anthropic) has a strong emphasis on nuanced, contextual communication. ChatGPT is optimized for task completion. Gemini, in writing tasks, still feels like it's optimizing for something closer to search relevance than human persuasion.
How to Use This Knowledge to Write Better Cold Emails Starting Today
Don't just pick the "winning" model and call it done. The smarter move is to use each tool for what it's actually best at — then combine the outputs in about 15 minutes.
Step 1: Use Claude to write your first draft. Give it a brief like the one above — trigger event, specific achievement with a number, target's role, word limit, and a single clear CTA. Prompt: "Write a cold email that leads with the reader's problem, not my credentials. Use [trigger event] as the opening hook. Include [achievement] as proof. Keep it under 150 words. Subject line included." Claude will almost always produce a problem-first draft that feels human.
Step 2: Paste Claude's draft into ChatGPT and ask it to stress-test the logic. Prompt: "Here's a cold email draft. Tell me: Does the subject line create enough curiosity? Is the CTA specific enough? Does anything read as vague or generic? Be direct." GPT-4o is excellent at analytical critique. It'll catch weak CTAs and vague language that Claude sometimes lets slide.
Step 3: Use ChatGPT to write two subject line variations. Prompt: "Write 3 subject line options for this cold email. One should be curiosity-driven, one should be outcome-driven, one should reference a trigger event. Keep each under 8 words." GPT-4o's output here tends to be sharper than Claude's for rapid-fire variation tasks.
Step 4: A/B test with real sends. Tools like Instantly or Lemlist let you split-test subject lines automatically. Send 50 emails with Claude's subject line, 50 with ChatGPT's best variation. You'll have real data within 72 hours.
The whole workflow takes 20 minutes once you've done it twice. And the email you end up with is genuinely better than anything any single model produces alone.
The Part Most People Get Wrong
Most people write their AI prompt like a job posting: "Write a cold email for a freelance designer looking for work." That's wrong. Vague input = vague output. Every time.
The models aren't mind-readers. When you don't give them a trigger event, a specific number, a target role, and a word limit, they fill the gaps with generic filler — and generic filler is exactly what gets you a 1% reply rate.
The second mistake: treating the first output as the final email. No professional copywriter sends their first draft. You shouldn't either. The AI's first response is a strong starting point, not a finished product. Run it through the critique step above, every single time.
The third mistake is the one I see most from people who've been doing this a while: using the same tool for every task in the workflow. ChatGPT isn't the best at everything. Claude isn't either. The people getting 15–20% cold email reply rates with AI are the ones using each model for its specific strength — not the ones who picked a favorite and stopped experimenting.
Key Takeaways
- Claude 3.5 Sonnet: Writes the most human-sounding, problem-first cold emails — use it for your first draft when emotional resonance matters.
- ChatGPT GPT-4o: Best for analytical critique, subject line variations, and task-specific rewriting — use it to pressure-test your draft.
- Gemini 1.5 Pro: Defaults to feature-benefit structure — not ideal for cold outreach, but still useful for longer-form content where comprehensiveness matters.
- Problem-first framing: Leading with the reader's pain before your credentials is the single biggest lever in cold email performance — always prompt for this explicitly.
- The multi-model workflow: Using Claude to draft, ChatGPT to critique, and real send data to validate is the approach getting the highest reply rates right now.
What to Do Right Now
Open Claude at claude.ai and paste this prompt: "Write a cold email from [your name/role] to [target role] at a [company type]. They recently [trigger event]. I have [specific achievement with a number]. Goal: get a 20-minute call. Keep it under 150 words. Lead with their problem, not my credentials. Include a subject line." Fill in your specifics, send the draft to ChatGPT for a critique, and you'll have a better cold email in the next 10 minutes than most people produce in an hour.