The best AI for writing cold emails, tested on 47,000 sends

Four tools, one prompt, one offer, one lead list, and about 47,000 sends in a week to settle it. If you want the answer to which is the best AI for writing cold emails without burning a week of sending capacity finding out: Claude Code won with 12 interested responses and a 42.86% interested rate. Gemini came second. Manus came last, and on rate it was not close.

This page is the written version of the test below. The ground rules that kept it honest, the full results by tool and by offer, the winning script sitting next to the losing one, and the prompting method that actually decided it. The tool matters less than people think. What you feed it matters more.

The three rules that kept it a real test

It is far too easy to keep re-prompting a tool until something eventually sounds good. That is not a test, that is you writing the email slowly. So the test ran on three non-negotiables:

1. Every tool gets the exact same prompt. Same website context, same two offers, same target: founders, CMOs, owners, and heads of growth at B2B marketing agencies.
2. Two re-prompts maximum, each. Any more than that and you could have written the emails yourself. Nobody running outreach at scale has half a day to spend typing prompts.
3. Same offer and same lead list across the board. The only variable was the tool itself.

The four winners went into the sequencer as eight angles, two offers per tool, split tested and launched at the same moment, pushing around 10,000 emails a day. The contenders were ChatGPT, Gemini, Claude Code (the agent, not the chat window, which is a real distinction here), and Manus, which behaves less like an LLM chat and more like an autonomous agent that breaks the task down on its own.

The two offers on trial

Offer one was a 2-minute Loom teardown of how the multi-channel system works. Offer two was a no-cost look-alike lead list: 100 companies that mirror their best client. Both are front-end offers, both are borderline free, and running both through all four tools meant the test could separate copy quality from offer quality instead of blending them into one number.

The results, by tool and by offer

Interested responses, not opens (we do not track opens on cold campaigns anywhere). Left to right: the Loom angle, the look-alike list angle, the combined total, and the interested rate on replies.

ChatGPT is the bar here, not because it is best, but because it is the default. It is the first thing most people open, so whatever it outputs is roughly what the market is sending right now. Two tools beat it on the same inputs.

The honest caveat: the week was messy. Reply rate across the account sat at 0.55%, and campaigns had to be split mid-test for deliverability reasons, which is something happening industry-wide as Google's bulk sender requirements keep tightening. It still produced over 73 interested responses in the last 7 days at roughly a 40% interested response rate, so the ranking is real even if the sending conditions were not clean.

Why Claude Code won, line by line

Its best performing script, in full:

If we could build you a no-cost list of 100 companies that look like {{case_study}}, same shape, same buying signals, would that be useful? PS, 24-hour turnaround, takes you 2 minutes to skim.

Four things are doing work in 30 words:

1. "No-cost" lands in the first six words. The reader knows what this costs them before they know what it is.
2. The case study variable, not the company name variable. Three of the four tools reached for company name by default. Every sender on earth uses that one, and it is often wrong anyway. Pulling their best case study off their site and building the list around it is a completely different email.
3. "Same shape, same buying signals" is elaboration, not padding. It tells them the list is built, not filtered.
4. The PS carries the close. A 24-hour turnaround signals someone is actually building this for them rather than blasting it. Two minutes to skim makes yes cheap.

Now the same offer from the tool that finished last, so the comparison is offer to offer:

Spending too much time prospecting for new high-value clients? We can generate a look-alike list of 100 companies matching the profile of your case study client, open to a 2-minute skim of the CSV.

The value is still in there. It is just buried behind a line that paints the picture first, and painting the picture reads like a cheesy ad to a stranger who has seen a hundred of them. They can be gone by the comma. Claude Code got to the value immediately; Manus got there in the second clause. On the same list, in the same week, that is the difference between first and last. The general principle behind that shape is broken down in how to write a cold email that gets replies.

The offer beat the tool in all four campaigns

Look at the table again by column instead of by row. The look-alike list beat the Loom teardown for every single tool: 8 to 4, 8 to 2, 7 to 2, 6 to 3. Not one exception.

The theory is straightforward. The list asks NOTHING of them and hands over free value with no back end attached. A Loom is a system you are showing them so you can install it later, so even when the video is useful, they can feel the ask waiting behind it. That, plus everyone gets Loom pitches all week and almost nobody gets offered a built list.

Which is the real takeaway from a tool bake-off: the spread between the best and worst offer was wider than the spread between the best and worst tool. Pick the tool second.

The prompting method that decided it

Three inputs, and they matter in this order:

1. A campaign brief before a single line of copy. Run the company URL through a research tool and ask for one thing: their ideal customers, geographic focus, persona segments by industry, job title and company size, the pain each segment carries, and how the company actually solves it. That last part is the one people skip and the one the copy lives on. The brief goes in as context, every time.
2. Project knowledge, meaning your own winners. Load your best performing past scripts and the frameworks behind them before you ask for anything. This is where an agent with file access separates from a chat window. Claude Code went to the website, read what the business does, and took visibly longer doing it. The faster tools answered faster and more generically, which is the same thing.
3. Re-prompt with direction, not dictation. Both re-prompts in this test were the same two notes: make the mechanism sound like something we built internally rather than a list of channels, and use the case study variable instead of the company name variable. Give it the idea. If you are writing the sentence, you did not need the tool.

Here is the part nobody selling AI copy will tell you. Even with the URL sitting in the context, not one of the four tools pulled a specific case study or value prop off the website on its own. They all had the raw material and none of them used it. AI drafts fast, it does not have taste, and you are still supplying the offer, the list, and the judgment call on which draft ships. The six jobs we actually hand it day to day are in 6 ways we use Claude Code for cold email, and if you want the wider tool landscape rather than a copy bake-off, that is the AI SDR tools breakdown.

tl;dr: Claude Code wrote the best cold email copy of the four, 12 interested at 42.86%, but the look-alike list offer beat the Loom offer for all four tools. Fix the offer, then feed the tool a real brief and your own winners.

FAQ

If you would rather not spend a week of sending capacity testing AI tools yourself, that is what we do: offer, lists, copy, infrastructure, follow-up, and 3,000+ sales calls booked for clients in the past year. Or keep learning free. Every training, grouped by topic, sits on the trainings hub, and the full video for this one is on YouTube.

PS - the winning email was 30 words and contained zero personalization beyond one variable. Four AI tools competed, and the thing that won was a good offer written plainly.