The best AI model for email
A reply you could have written, shorter and calmer.
Short answer
ChatGPT Free ($0) and GPT-5.6 Luna are the subscription fit. On the API, DeepSeek V4 Flash 0423 at $0.09 in · $0.18 out per million tokens The rate is for 1,000,000 tokens, not for one message. A short email is a tiny fraction of that. A long contract with a long answer can use a real slice of it. Glossary is the cheapest sensible pick. This illustration A chunk of text the bill is counted in. A token is often a short word or part of a word. One million tokens is roughly 750,000 English words. That ratio is a rule of thumb, not a count of your file. Glossary is 800 input tokens and 350 output tokens, about $0.0001.
Daily email is a reply, a nudge, or a recap of a thread you paste. The model matches the tone you ask for. You delete anything too smooth. Do not ask it to remember a promise that is not in the thread.
Subscription
ChatGPT Free’s Instant mode is GPT-5.6 Luna, the light model on that plan. A short prompt you can paste fits this lane. Plus is the first plan that adds GPT-6 Sol. The dollar amount for Plus was not in the pricing-page text captured on 20 Nov 2026, so this guide does not invent one.
ChatGPT plans · Check a plan you already pay for
API
A thread you paste is small. DeepSeek V4 Flash is the cheap pick. The illustration is a short thread in and a reply out.
DeepSeek V4 Flash 0423 lists a context window How much text the model can hold in one request: your prompt, the files, and the reply together. If the document is bigger than the window, it does not fit in one go. Glossary of 1,048,576 tokens. A short everyday paste uses a sliver of that.
| Lane Quality, balanced, and cheapest sensible are editorial lanes, not a lab’s product names. Quality is the reviewed pick for hard work. Cheapest sensible is the reviewed pick when a miss is easy to catch. Glossary | Model | Rate What you pay for tokens you send. $2 in means $2 per million tokens of your prompt, files, and images. A rough English guide is about 750,000 words per million tokens. Your file can be shorter or longer. This is not a measurement of a specific document. Glossary | This illustration What you pay for tokens the model writes back. $10 out means $10 per million tokens of the reply. Long answers cost more than short ones, even when the question was small. Glossary |
|---|---|---|---|
| Cheapest sensible | DeepSeek V4 Flash 0423 | $0.09 in · $0.18 out | $0.0001 |
| Step up | GPT-6.2 Sol | $2.00 in · $10.00 out | $0.0051 |
Input is what you send What you pay for tokens you send. $2 in means $2 per million tokens of your prompt, files, and images. A rough English guide is about 750,000 words per million tokens. Your file can be shorter or longer. This is not a measurement of a specific document. Glossary. Output is what the model writes back What you pay for tokens the model writes back. $10 out means $10 per million tokens of the reply. Long answers cost more than short ones, even when the question was small. Glossary. $0.09 in and $0.18 out are per million tokens, not per message.
Step up when
- The thread is angry, legal, or going to many people. Read a Flash draft. If the tone is off in a way that matters, try Sol on that one message, not on every email.
- The input is a photo, a scan, or a file rather than words you can paste. GPT-6 Luna lists file and image input on the feed. DeepSeek V4 Flash lists text only.
- You need the model to know what was decided last month and you do not have the mail in front of you. Paste it. A larger model does not have your inbox.
The step-up model on this page is GPT-6.2 Sol, $2.00 in · $10.00 out per million tokens. The same illustration is about $0.0051.
Try these prompts
The late reply
Draft a short reply. Apologize for the delay, answer the two questions in the thread, and propose Thursday at 2. Do not add other commitments.
The firm no
Say no to this request in under 80 words. Stay polite. Give the reason I wrote below and no extra excuse.
The thread recap
Summarize this pasted thread into decisions, open questions, and who does what. If the thread does not name an owner, write “not named.”
Overkill
In the app you already use
Only features with a vendor page. Checked 20 Nov 2026.
- Copilot and Outlook — Microsoft’s Copilot page says Copilot is integrated into Outlook. The personal-plans support page says you can summarize an email discussion in Outlook. Those pages do not name the model or a message quota.
- Claude can take the export — Claude’s help center says you can upload documents, including text and PDF, into a chat. A mailbox export you are allowed to share can be one of those files. The article publishes size limits. It does not say Claude is connected to your mailbox unless you add that yourself.
Questions
Will it sound like me?
Only if you paste a few of your own sentences and say to copy that tone. Otherwise it sounds like a careful stranger.
Should I use deep research to answer email?
No. A reply is the mail in front of you. Deep research is a different, slower feature for when you have to go find sources. It is described on the research page.