Data · published 6 November 2026
What it costs to have AI write 1,000 email replies
As of 6 November 2026, at list prices, having AI write 1,000 email replies costs between $0 on a model that runs on your own Mac and about $26.88 on the most expensive hosted model we priced (Claude claude-opus-4-8, for 1,000 long-thread replies). For 1,000 short replies the range is $0.20 (Groq openai/gpt-oss-120b) to $7.31 (Claude claude-opus-4-8); for 1,000 typical replies written in your own style it is $0.41 to $14.81.
Cost per 1,000 replies
| Provider and model | Short reply | Typical reply, style learning | Long thread |
|---|---|---|---|
| Groqopenai/gpt-oss-120b | $0.20 | $0.41 | $0.75 |
| Anthropic Claudeclaude-opus-4-8 | $7.31 | $14.81 | $26.88 |
| OpenAIgpt-5.4 | $4.03 | $7.97 | $14.38 |
| xAI Grokgrok-4.3 | $1.27 | $2.86 | $5.31 |
| Google Gemini (to 2026-12-31)gemini-3.8-flash | $1.10 | $2.22 | $4.03 |
| Google Gemini (from 2027-01-01)gemini-3.8-flash | $2.19 | $4.44 | $8.06 |
| DeepSeek (peak, cache miss)deepseek-flash | $0.39 | $0.82 | $1.50 |
| DeepSeek (off-peak)deepseek-flash | $0.20 | $0.41 | $0.75 |
| Apple Intelligenceon-device | $0.00 | $0.00 | $0.00 |
| Ollamalocal | $0.00 | $0.00 | $0.00 |
| Local OpenAI-compatible server (LM Studio, llama.cpp)local | $0.00 | $0.00 | $0.00 |
Cost per single reply
| Provider | In / out $ per M tokens | Short | Typical | Long thread |
|---|---|---|---|---|
| Groq | $0.15 / $0.60 | $0.0002 | $0.0004 | $0.0008 |
| Anthropic Claude | $5 / $25 | $0.0073 | $0.0148 | $0.0269 |
| OpenAI | $2.50 / $15 | $0.0040 | $0.0080 | $0.0144 |
| xAI Grok | $1.25 / $2.50 | $0.0013 | $0.0029 | $0.0053 |
| Google Gemini (to 2026-12-31) | $0.75 / $3.75 | $0.0011 | $0.0022 | $0.0040 |
| Google Gemini (from 2027-01-01) | $1.50 / $7.50 | $0.0022 | $0.0044 | $0.0081 |
| DeepSeek (peak, cache miss) | $0.30 / $1.20 | $0.0004 | $0.0008 | $0.0015 |
| DeepSeek (off-peak) | $0.15 / $0.60 | $0.0002 | $0.0004 | $0.0008 |
| Apple Intelligence | $0 / $0 | $0.0000 | $0.0000 | $0.0000 |
| Ollama | $0 / $0 | $0.0000 | $0.0000 | $0.0000 |
| Local OpenAI-compatible server (LM Studio, llama.cpp) | $0 / $0 | $0.0000 | $0.0000 | $0.0000 |
Method
We modelled three reply sizes. Token counts are estimated as characters divided by four, the same estimate MailTwin's own cost ledger uses; real tokenizers differ by model and language.
- Short reply: incoming email 1,200 characters, your notes 150, system prompt 1,500, which is 2,850 characters in (712.5 tokens) and 600 characters out (150 tokens).
- Typical reply with style learning: the same email and notes, but the system prompt carries three to five of your own sent emails as style samples, about 6,000 characters in total. Input is 7,350 characters (1,837.5 tokens); output is 900 characters (225 tokens).
- Long thread: incoming thread 8,000 characters and system prompt 6,000, which is 14,000 characters in (3,500 tokens), with no separate notes; output is 1,500 characters (375 tokens).
Cost per reply = input tokens / 1,000,000 x input price + output tokens / 1,000,000 x output price. The per-1,000 figure is that value multiplied by 1,000, rounded to the cent. The figures were computed with a short script, not by hand.
Prices are per million tokens, list price, verified against the providers' catalogs on 13 October 2026: Groq openai/gpt-oss-120b $0.15 in / $0.60 out (free tier of 30 requests per minute and 1,000 per day); Claude claude-opus-4-8 $5 / $25; OpenAI gpt-5.4 $2.50 / $15; Grok grok-4.3 $1.25 / $2.50; Gemini gemini-3.8-flash $0.75 / $3.75 through 31 December 2026 and $1.50 / $7.50 from 1 January 2027 (a free tier exists); DeepSeek deepseek-flash $0.30 / $1.20 at peak with a cache miss, half that off-peak. Apple Intelligence, Ollama and local OpenAI-compatible servers cost $0 per request because they run on your Mac. Sources are the providers' own pricing pages: Anthropic (anthropic.com/pricing), OpenAI (openai.com/api/pricing), Google AI (ai.google.dev/pricing), xAI (docs.x.ai), Groq (groq.com/pricing) and DeepSeek (api-docs.deepseek.com).
What this leaves out
- Cached-input discounts, which can lower the cost of a repeated system prompt considerably.
- Batch-API discounts, which do not apply to interactive replies but matter for bulk work.
- Free tiers, which can bring light use to $0 on Groq and Gemini within their rate limits.
- Reasoning tokens, which some models bill as output and which can raise the output side well above the visible reply length.
- Taxes and currency conversion, and the electricity and hardware cost of running a local model.
These are estimates from list prices. The provider's own billing is authoritative.
Why we worked this out
MailTwin is a Mac app that adds AI replies to Apple Mail using your own key with any of these providers, or a model on your Mac. MailTwin
Cite this
MailTwin, What it costs to have AI write 1,000 email replies, 6 November 2026, https://www.mailtwin.ai/ai-email-reply-cost