We Gave ChatGPT, Claude, and Gemini the Same Cold Email Brief — Here's What Happened
One wrote something you'd actually want to send. One produced a wall of corporate fluff that would've gone straight to spam. And one — I'm not joking — included a line so legally questionable that a real sales rep could get in serious trouble for sending it. We ran a head-to-head test using the exact same brief across all three AI tools, and the differences were sharper than most people expect. By the end of this article, you'll know which AI to use for cold outreach, why the other two fall short in specific ways, and how to prompt each one to get the best possible result. Let's start with what we actually asked them to write.
The Brief Was Identical — The Results Were Anything But
The test brief was deliberately realistic. Here's exactly what we gave all three AIs:
"Write a cold email from a B2B SaaS company called Clearpath to a VP of Operations at a mid-sized logistics company. The goal is to book a 15-minute demo. Clearpath helps companies reduce delivery delays by 30% using AI-powered route optimization. Keep it under 150 words. Don't use buzzwords. Make it feel human."
Same prompt. Same word count limit. Same request to avoid buzzwords. Three very different outputs.
ChatGPT (GPT-4o) delivered a clean, punchy email with a specific subject line, a two-sentence value pitch, and a soft CTA that didn't feel pushy. It clocked in at 127 words, respected the brief, and most importantly — it sounded like a real person wrote it. The subject line alone ("Your Q4 delivery windows — quick thought") was good enough to steal.
Claude (claude-3.5 Sonnet) went a layer deeper. It opened with a pain-point observation specific to logistics ("Q4 is brutal for last-mile delays") and then connected Clearpath's solution directly to that moment. The tone was warmer, almost conversational. It also gave you a second version unprompted — shorter and more direct — which was a nice touch. Claude's instinct to show empathy before pitching is something most human copywriters don't even do consistently.
Gemini is where things got uncomfortable. The email it produced was technically correct but read like a marketing brochure collapsed into an email format. Phrases like "industry-leading AI solutions" and "cutting-edge optimization technology" appeared despite explicitly asking for no buzzwords. Then there was this line: "Our platform has been proven to eliminate delivery failures entirely." That claim is not only unrealistic — it's the kind of absolute guarantee that opens you up to legal exposure if a prospect ever screenshotted it and held you to it.
Why Claude Wins the Nuance War (And When That Actually Matters)
Most cold email comparisons stop at "which one sounded best." That's the wrong question. The right question is: which AI understands the psychology of cold outreach?
Cold emails fail for one reason more than any other — they feel like they're about the sender, not the recipient. Claude consistently flips this. In our test, it structured the email recipient-first: problem acknowledgment, then relevance, then ask. That's a framework professional copywriters call Problem-Agitate-Solve, and Claude applied it without being told to.
Here's the deeper thing most people miss: Claude is trained with a strong emphasis on helpfulness and avoiding harm, which in writing terms translates to restraint. It doesn't oversell. It doesn't make claims it can't back up. For cold email — where trust is everything and you have about four seconds to earn it — that restraint is a superpower.
Try this prompt with Claude to see what we mean: "Write a cold email for [your product] targeting [specific role]. Open with a one-sentence observation about a challenge they face in Q4. Don't pitch the product until sentence three. End with a question, not a CTA button." The output will feel different from anything ChatGPT produces by default.
That said, Claude's warmth can tip into too casual if you're selling into a formal industry like finance or legal. If your prospect is a CFO at a bank, you probably want less "Hey Sarah" and more precision. That's when ChatGPT's more neutral tone becomes the better pick.
How to Use All Three Together in One 20-Minute Workflow
Here's something almost no one talks about: you don't have to pick one. The smartest cold email workflow uses all three AIs in sequence, each doing what it's best at.
Step 1 — Use Gemini for research (5 minutes). Gemini has live web access, which means it can pull recent information about your prospect's company. Start with: "Summarize the top 3 operational challenges facing mid-sized logistics companies in Q4 2024 based on recent news and industry reports." Don't use Gemini to write the email. Use it to brief yourself before you write anything.
Step 2 — Use Claude for the first draft (10 minutes). Take the context from Gemini and feed it to Claude with your core brief. Prompt: "Here's background on the prospect's industry challenges: [paste Gemini output]. Now write a cold email from [your company] to a VP of Operations. Lead with their pain, not our product. Max 130 words." Claude's empathy-forward writing style shines when it has real context to work with.
Step 3 — Use ChatGPT to tighten and test subject lines (5 minutes). Paste Claude's draft into ChatGPT and prompt: "Tighten this cold email to under 120 words without losing any specificity. Then write 5 subject line options — two curiosity-based, two direct, one that uses their company name." ChatGPT is excellent at mechanical refinement and variation generation. It's less inspired than Claude but more consistent as an editor.
In 20 minutes you've used AI-powered research, psychologically smart drafting, and precise editing. That's a cold email workflow most agencies charge four figures to build out.
The Part Most People Get Wrong
Most people use one AI, hit send on the first output, and wonder why their reply rate is 2%.
The problem isn't the AI — it's the prompt. Generic brief in, generic email out. If you don't tell the AI who the recipient is, what specific problem they have, and what a realistic ask looks like, you'll get something that sounds like every other cold email in their inbox. The AI is only as specific as the information you give it.
There's a second mistake that's just as common: using the AI's output as a final product instead of a first draft. Every email these tools produce needs a human pass. Read it out loud. Does it sound like you? Does it reference something real about the recipient's world? If not, it needs one more revision — and that revision should take you two minutes, not twenty.
The third mistake is optimizing for length when you should be optimizing for clarity. People obsess over "keep it under 100 words" as if word count is what earns replies. What earns replies is one clear idea, one credible claim, and one obvious next step. You can do that in 80 words or 160 words. Stop counting words and start counting ideas — there should only be one.
Key Takeaways
- ChatGPT (GPT-4o): Best for clean, neutral-toned emails and fast iteration on subject lines — especially when selling into formal or technical industries.
- Claude: Best for emotionally intelligent, empathy-first drafts that don't oversell — strongest when you feed it real context about the prospect's challenges.
- Gemini: Best used for pre-writing research using its live web access — not recommended as your primary cold email drafting tool.
- The three-AI workflow: Using Gemini for research, Claude for drafting, and ChatGPT for editing gives you better output than any single tool alone — in under 20 minutes.
- Prompt specificity: The quality of your cold email output is almost entirely determined by how much relevant context you put into your prompt — vague brief, vague email.
What to Do Right Now
Open Claude right now and paste this prompt: "Write a cold email from [your company] to a [specific job title] at a [specific industry] company. Their biggest challenge right now is [one specific pain point]. Our product helps with that by [one specific mechanism]. Max 130 words. Lead with their problem, not our product. End with a question." Fill in the brackets with your real details, send the output through one human read-aloud pass, and you'll have something worth testing by the end of today.