I tested all available versions of ChatGPT to see which model writes the best work emails

  • OpenAI unveiled GPT-4o, its latest generative AI model, on Monday.

  • GPT-4o is available as a text-based chatbot on ChatGPT.

  • Business Insider tested GPT-4o's email-writing capabilities and compared it to older models.

OpenAI's newest model, GPT-4o, is predicted to shake up multiple industries, but its email writing skills might need some work.

During a flashy display Monday, the company unveiled the latest version of the generative AI model predicted to shake up multiple industries: GPT-4o.

GPT-4o is already available as a chatbot on ChatGPT, which many workers use for boring, text-based tasks. Free ChatGPT users are stuck with GPT-3.5 but get limited access to a handful of responses from GPT-4o, per OpenAI's pricing page.

The ChatGPT Plus membership costs $20 a month and includes more answers from the faster GPT-4o model, access to GPT-4, and features like an image generator.

As someone who has previously tried and grown frustrated with ChatGPT's inability to write non-robotic-sounding emails, I wanted to see if GPT-4o would actually help me out for once.

Using ChatGPT's "temporary chat" feature to prevent the chatbot from spitting out similar responses based on previous answers, I tested three available versions of the chatbot — GPT-3.5, GPT-4, and GPT-4o — using three different email prompts to see if GPT-4o is superior to the earlier versions. Here's what I found.

Interview requests

A text recreation of a conversation between the author and GPT 4o. The author is asking the chatbot to write an email about the history of generative AI.
ChatGPT's newest model gave the shortest answer for this prompt.Business Insider

This is one of the most common emails I send at my job, so I was an expert on how they should be structured. I've tried having GPT write these emails, but it never works out — the emails ChatGPT constructs are always extremely long and ridiculously official-sounding. I'm sending an email, not a quill-and-inkpot letter to the King of England.

I had hoped GPT-4o would change its ways, but it's hard to teach an old chatbot new answers.

Here's the prompt I gave:

Can you write an email to an expert I am interested in interviewing about the history of generative AI models? I want the email to be brief and friendly. Let the expert know I work at Business Insider.

GPT-4o returned the shortest draft, which I was grateful for. However, I did find it odd that the chatbot wrote that I would be "honored" to interview a subject. Do not worry, future interview subjects — I will not come off this strong in an email.

The older models provided lengthier answers. I did find it interesting that these models made up topics my fake source and I could discuss, and that they were topical. For example, the older GPT-4 model said the faux interview would "focus on the key milestones and future implications of generative AI technology."

GPT-4o's answer wasn't quite as detailed — it took the "brief" part of my request seriously.

Winner: GPT-4o

Scheduling a meeting

A text recreation of a conversation between the author and GPT 4o. The author is asking the chatbot to write an email to set up a virtual meeting with a boss.
All three models gave satisfactory responses to this prompt.Business Insider

Sending meeting emails isn't particularly difficult, but it's always easier to have someone — or something — else write them for you.

My second prompt:

I want to schedule a virtual meeting with my boss to discuss an ongoing project. Can you write an email asking to put something on the calendar this month? I want the email to be brief and conversational.

I was surprised that all three models constructed their emails concisely and similarly. My simple and straightforward prompt may have helped.

I found it interesting that the previous generation, GPT-4. asked me to elaborate on the meeting topic at one point to draft a better email. Still, all three models gave responses that would work well for someone trying to communicate with their higher-ups.

Winner: Three-way tie

Cold emailing about a career opportunity

A text recreation of a conversation between the author and GPT 4o. The author asks the chatbot to write an email introducing them to someone at JPMorgan.
All the answers were lengthy — and similar.Business Insider

A cold email: It's probably one of the scariest messages that someone can send. For my final prompt, I asked ChatGPT to take on the anxiety-inducing task for a persona I made up — a financial advisor looking for a gig at JPMorgan.

I am a budding financial advisor looking for a new job. I would like to introduce myself by cold-emailing someone at JPMorgan. Can you write this email for me and include details that would make me stand out?

For this one, I did not add a line about brevity, resulting in all of the models churning out responses that were far too long. Interestingly, GPT-4o wrote the longest email, at nearly 300 words. The other two were around 250.

All the emails were structured very similarly. For example, all three lost sight of the topic and thanked the email recipient for "considering my application."

Both the GPT-4o model and the 3.5 model added the phrase "What sets me apart…" which indicated to me that it was pulling from my prompt. It was something I would expect to read in a middle school essay — not from an adult job seeker.

Winner: GPT-4

And the winner is...

ChatGPT logo repeated three times
ChatGPT, Tyler Le/BI

ChatGPT will be ChatGPT.

The bot has a distinct voice and style that makes it obvious that a non-human is composing whatever text output you ask for.

I don't think the newer model has shed that sound on paper — though maybe having the text read aloud by a human-sounding, and perhaps flirty, voice assistant, as OpenAI recently demoed, would change things.

The text-generating capabilities of each model were fairly similar to me. Sometimes, the older models gave me answers that I found more conversational and less robotic. So, if you only want ChatGPT to write emails for you, the monthly $20 is probably not worth it.

However, if you want to avoid sending emails and generate images of cute cats in Amazon packages to pass the time at your desk, it may be time to pull out your wallet.

OpenAI did not immediately respond to a request for comment.

Read the original article on Business Insider