This post is a first-person take, not a benchmark. I’ve been writing 30–40 work emails a day for 8 months now (mix of internal team comms, client-facing escalations, and the occasional “no, we’re not extending the deadline” pushback). I tried ChatGPT, Gemini, and Claude on the same emails for 3 months. Eventually I stopped reaching for the others.

Here’s the comparison, what each one is actually good at, and why “best AI” is the wrong question.

The trial setup

For 12 weeks, every time I had a non-trivial email to write, I’d:

  1. Draft a bullet list of points to make.
  2. Send the same prompt to ChatGPT (4o), Gemini (2.5 Pro), and Claude (Sonnet 4.5).
  3. Pick the best output.
  4. Track which one I picked.

By week 8, the count was: Claude 64%, ChatGPT 26%, Gemini 10%.

By week 12, I’d stopped using Gemini entirely. ChatGPT was still in the rotation for one specific use case (more on this). For everything else, Claude.

What Claude does better

Specifics. Claude handles ambiguous requests by inferring the most likely scenario and writing for it. Example prompt I used multiple times:

“Write a polite escalation email to the client account manager about the stuck contract review. We’ve been waiting 3 weeks. Don’t sound annoyed, but make it clear we need a date.”

Claude’s draft felt close to how I’d write it: opened with context, named the specific blocker (“legal review of clause 4.2”), proposed a date (“can you commit to Friday this week?”), closed warmly without being saccharine.

ChatGPT’s draft was generic. “I wanted to follow up on our pending discussion.” “Looking forward to your prompt response.” Useful as a structure, but I had to rewrite every paragraph for voice.

Gemini’s draft was passable but had this corporate-newsletter feel — overly formal, with phrases like “I trust this email finds you well” that haven’t been used by a human in 30 years.

Tone control. When I asked for “polite but firm” or “direct but not aggressive,” Claude understood the gradient. ChatGPT often went too soft (“I just wanted to gently mention…”); Gemini stayed in the corporate-template register regardless.

Editing existing drafts. This is where Claude is most ahead. If I write a rough email and ask “tighten this, remove the parts that sound defensive,” Claude actually does it. ChatGPT sometimes rewrites the whole thing, which I have to undo. Gemini occasionally misreads “tighten” as “expand.”

Following meta-instructions. “Don’t use em dashes.” “Avoid the word ‘leverage’.” “Sign off with ‘Thanks’ not ‘Best regards’.” Claude remembers these across the conversation. ChatGPT forgets after 3-4 exchanges. Gemini occasionally ignores them entirely.

What ChatGPT does better

Two things, genuinely:

1. Tabular and bulleted formatting. When I need an email with a comparison table or a structured list (for example, summarising 5 vendor quotes), ChatGPT’s markdown rendering is better. Claude defaults to prose; ChatGPT defaults to tables. For “show me the data” emails, ChatGPT wins.

2. Multilingual edge cases. I occasionally write emails in Hindi, Hinglish, or Marathi for vendor coordination. ChatGPT’s multilingual handling is more natural — Claude does it well in English-Hindi mix, but a pure-Hindi email from Claude sometimes reads like translated English. ChatGPT (especially 4o and later) writes Hindi that feels like Hindi.

For these two cases, I still keep ChatGPT around.

Where Gemini falls short

Gemini was the easiest decision. Two recurring problems:

  1. Tone dial doesn’t go warm. Gemini’s professional emails always feel like they were written by HR Compliance. The default voice is reserved, formal, and slightly defensive.
  2. Awkward “I” usage. Gemini frequently writes “I will be sending the report tomorrow” instead of “I’ll send the report tomorrow.” Small thing; adds up across 40 emails.

These are voice issues, not capability issues. Gemini knows the right facts; the writing just doesn’t sound human enough.

My actual email workflow now

Here’s the loop I run for every non-trivial email:

  1. Bullet the points in a Notion page or a quick text file. 30 seconds of thinking — the actual content of what to say.
  2. Send to Claude with the bullets, target tone, and any meta-rules (“no em dashes, sign-off as ‘Thanks’”).
  3. Read the draft. 60% of the time, it’s sendable as-is. 30% of the time, I edit one or two sentences. 10% of the time, I rewrite from scratch (usually because I missed a key bullet).
  4. Final pass for voice. Replace anything that sounds AI-written. Common fixes:
    • “I hope this email finds you well” → delete entirely
    • “I would like to” → “I’d like to” (or just delete)
    • “Please feel free to” → “Let me know if”
    • “I appreciate your understanding” → “Thanks for the heads up” (or just delete)

The whole loop takes 2–3 minutes for an email I’d have spent 8 minutes writing from scratch.

Things AI still doesn’t do well for emails

Reading the room. If a colleague is in a tense thread with a client, the AI doesn’t know that. It writes the “ideal” email; you have to inject the relational context.

Internal politics. The email to your boss who has explicitly told you not to escalate things to the VP — the AI will happily draft an email that copies the VP because it doesn’t know your political situation.

Speed for trivial replies. “OK”, “Thanks”, “Got it, will do by EOD” — for these, typing is faster than prompting. Use AI for emails over 100 words. Below that, it’s overhead.

Multi-thread context. If a thread has 12 messages, you’d be feeding all of them to the AI for context. Claude handles this fine but it’s annoying. Gmail’s “smart compose” is better for in-thread replies because it’s already inside Gmail.

Companion tool

Our Email Writer packages this workflow — bullet-point input, tone selection, output you can copy. Behind the scenes, the tool uses Claude with a system prompt that bans the same set of cliches I avoid manually (“hope this email finds you well”, “kindly”, “feel free to”, “looking forward”).

If you’re testing tools yourself, run the same prompt through ChatGPT, Claude, and Gemini for a week. Pick what wins for your writing voice. The “best” AI isn’t a fixed answer — Claude wins for me because I write in a particular register; someone who writes corporate-formal emails might prefer ChatGPT. The trial run takes 30 minutes a day for a week and saves you months of suboptimal tool choice.

TL;DR

  • Claude for everything that needs voice, tone, or editing.
  • ChatGPT for emails with tables, charts, or non-English content.
  • Gemini for nothing in this category, in my testing.

Try our Email Writer on your next sticky email and see how it lands.