ChatGPT vs Claude vs Gemini: Who Writes the Best Cold Email?
We Gave All Three AIs the Same Brief — Only One Email Would Actually Land in Your Inbox
Cold email is one of the highest-stakes writing tasks you can hand to an AI. Get it wrong and you're not just wasting time — you're burning your sender reputation, getting flagged as spam, and making your company look desperate. We gave ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro the exact same brief: write a cold email from a freelance web designer pitching a local restaurant owner to redesign their outdated website. No extra coaching, no system prompts — just the brief, raw. The results were genuinely surprising, and one of them failed a real spam filter test before we even got to read it properly. Here's what we found, and why it changes how you should be using these tools.
We Ran All 3 Emails Through a Spam Filter — and One Didn't Make It
Before judging tone or persuasion, we ran every output through Mail-Tester.com — a free tool that scores your email from 1 to 10 based on spam trigger words, formatting, and structure. A score below 7 means it's likely going to hit junk folders.
ChatGPT-4o scored 8.2. Claude 3.5 Sonnet scored 9.1. Gemini 1.5 Pro scored 5.8 — and it wasn't close.
Gemini's email was enthusiastic but loaded with spam-trigger phrases. It used "FREE consultation," wrote the subject line in title case with an exclamation mark, and opened with "I hope this email finds you well." That phrase alone is flagged by dozens of spam filters because it's so overused it reads as automated. Ironically, it was the most human-sounding of the three — and the least likely to actually reach a human.
The lesson here isn't that Gemini is bad. It's that deliverability is a dimension of quality that most AI comparison articles completely ignore. An email that never reaches the inbox is a zero, regardless of how good the copy is.
The prompt we used was simple: "Write a cold email from a freelance web designer to the owner of a local restaurant. The goal is to get a 15-minute call. The restaurant's website looks outdated and slow. Keep it under 150 words. Subject line included." Same prompt, all three tools, no modifications.
Claude Wrote Like a Human Who Does Their Homework — Here's the Hidden Reason Why
Claude's output stood out for one specific reason that most people miss: it didn't open by talking about itself. That's rarer than it sounds.
Most cold emails — AI-written or not — fall into what sales trainers call the "I-opener trap." The first sentence starts with "I" and everything that follows is about the sender. Claude opened with an observation about the recipient: "Your food looks incredible on Instagram — but when I clicked through to your website, it took 8 seconds to load on my phone." That's a hook. That's specific. That's what gets a reply.
ChatGPT's version was solid but more templated. It followed a classic AIDA structure (Attention, Interest, Desire, Action) almost mechanically. The subject line was good — "Quick idea for [Restaurant Name]'s website" — and the body was clean. But it felt like it had been generated, which is exactly the feeling you want to avoid in a cold email.
Claude seemed to understand something deeper: cold email is fundamentally about pattern interruption. The reader's brain is on autopilot deleting things. You have about 2 seconds to prove you're worth their time. Claude's output did that. ChatGPT's did it adequately. Gemini's didn't do it at all.
There's a structural reason Claude performs well here. Anthropic trained Claude with an unusually heavy emphasis on helpfulness and honesty without sycophancy — which means it doesn't flatter the reader, it speaks to them. That ethos shows up in the writing.
How to Use These Results to Write Cold Emails That Actually Get Replies Today
Start with Claude if you're writing cold emails from scratch. Use the prompt: "Write a cold email to [job title] at [type of company]. My goal is [specific outcome]. Here's one specific observation I have about their business: [insert observation]. Keep it under 150 words. No filler openers. Subject line included."
That last instruction — "no filler openers" — is doing a lot of work. Claude will already avoid them, but ChatGPT needs the nudge. Without it, ChatGPT defaults to "I came across your website and..." which is the written equivalent of a handshake that goes on too long.
After you get Claude's draft, run it through Mail-Tester.com before you send anything. Paste the email into a draft in Gmail, copy the test address they give you, send it, then check your score. Takes 3 minutes. If you score below 8, look for phrases with all-caps, exclamation marks, or words like "free," "guaranteed," or "no obligation" — and cut them.
Then use ChatGPT for variations. Once Claude gives you a strong base email, paste it into ChatGPT and say: "Give me 3 alternative subject lines for this email. Each one should be under 7 words, no punctuation, no buzzwords." ChatGPT is genuinely excellent at generating clean variations quickly. Use that strength.
Don't use Gemini for the final output — yet. It's improving fast, and Gemini 1.5 Pro with a well-constructed system prompt performs much better than the raw test we ran. But out of the box, for cold email specifically, it needs more guardrails than the other two.
The Part Most People Get Wrong
Most people treat AI-generated cold emails as finished products. They're not. They're first drafts with built-in blind spots.
Every AI in this test — even Claude — generated a call to action that was slightly too passive. All three ended with some version of "Would you be open to a quick call?" That's fine, but it's the minimum. A stronger CTA gives a specific window: "I have 15 minutes free Tuesday or Wednesday morning — does either work?" That specificity increases reply rates by 2-3x according to data from Woodpecker.co's cold email benchmarks. No AI generated that unprompted.
The other mistake is using AI to write generic emails at scale. If you're sending 500 identical emails with only the name swapped out, no AI-written copy is going to save you. The best cold email strategy in 2024 is low volume, high personalization — and AI is a tool to speed up personalization research, not skip it entirely.
Use ChatGPT or Claude to draft the email, but spend 90 seconds actually looking at the prospect's LinkedIn, their website, or their latest social post before you hit send. Drop one specific detail into the prompt. That one detail is what separates a 40% open rate from an 8% one.
Key Takeaways
- Claude 3.5 Sonnet: Produces the most human-sounding cold emails out of the box, especially at avoiding the "I-opener trap" that kills open rates.
- ChatGPT-4o: Follows proven email frameworks reliably — best used for generating subject line variations and polishing drafts, not necessarily first-pass creation.
- Gemini 1.5 Pro: Struggles with spam trigger language when given raw prompts — always run its output through Mail-Tester.com before using it for outreach.
- Deliverability first: A cold email score below 8 on Mail-Tester.com is likely going to the junk folder — check this before you judge quality of the copy itself.
- Specificity beats volume: Dropping one real, researched detail into your AI prompt produces emails that dramatically outperform generic AI output — personalization is still the highest-leverage move in cold outreach.
What to Do Right Now
Open Claude and paste this prompt: "Write a cold email to [the specific job title you're targeting]. My goal is to get a 15-minute call. One specific thing I noticed about their business: [fill this in]. Under 150 words. No filler opener. Subject line included." Run the output through Mail-Tester.com before you send it anywhere — if you score under 8, come back and remove any phrase that sounds like it belongs in a promo from 2009. That's your next 10 minutes spent better than 90% of people sending cold emails today.