AI email comparison
DeepSeek vs Perplexity for Email Drafting
You already use AI to draft emails. The question is which tool sounds most like you, and whether there's a better option than either.
DeepSeek
The Chinese open-source model that shocked the AI world with GPT-4 level performance at a fraction of the cost. Popular with BYOK users watching API pricing.
DeepSeek is an AI research lab that released a series of open-weight models, most notably DeepSeek-V3 (their flagship chat model), DeepSeek-R1 (a reasoning model competitive with o1), and DeepSeek-Coder-V2 (specialized for code generation tasks). The models use an OpenAI-compatible API format, which means any tool that accepts a custom endpoint and API key can route to DeepSeek with minimal configuration changes. For email drafting specifically, DeepSeek-V3 (accessed via the model name deepseek-chat) handles instruction-following and tone tasks well, though its training data skews toward Chinese-language text which can affect idiomatic English phrasing in subtle ways. The primary draw for BYOK users is cost: DeepSeek's API pricing is roughly 10-20x cheaper than GPT-4o at equivalent quality tiers, making it an attractive choice for high-volume drafting. The main concern for enterprise and regulated users is data residency: DeepSeek's inference servers are located in China, which creates compliance issues for organizations with strict data sovereignty requirements. Users who want DeepSeek's cost profile without the data residency risk typically run self-hosted versions via Ollama or a private cloud deployment.
Strengths for email
- Dramatically cheaper API pricing than OpenAI: roughly 10-20x less per token at comparable quality
- DeepSeek-V3 (deepseek-chat) is competitive with GPT-4o on instruction-following and email tasks
- DeepSeek-R1 (deepseek-reasoner) handles complex, multi-step reasoning comparable to OpenAI o1
- DeepSeek-Coder-V2 excels at code generation tasks
- OpenAI-compatible API format: works with any tool accepting a custom endpoint and one API key
- Open-source model weights available for self-hosting via Ollama or private deployment
- Excellent cost-performance ratio for BYOK users running high email volume
Weaknesses for email
- Data residency: inference servers are in China, a compliance risk for enterprise and regulated industries
- Training data skews toward Chinese-language text; subtle idiomatic gaps in English email prose
- Slow responses at peak load on the shared API; self-hosting resolves this but adds setup
- Less community support for email-specific prompting vs. the OpenAI ecosystem
- No native email client integration: requires copy-paste or a BYOK tool like ForthWrite
- Model weights are large; self-hosting requires significant hardware (GPU with 40GB+ VRAM for full precision)
Pricing: API: ~$0.14/M input tokens (DeepSeek-V3), ~10-20x cheaper than GPT-4o. Self-hosted: free, hardware costs only.
Best for: Cost-conscious BYOK users who want GPT-4 quality at a fraction of the API cost and can accept China-based data residency
Perplexity
An AI search engine, not primarily a writing tool. Excellent for research, but not designed for email drafting or voice matching.
Strengths for email
- Unmatched for research and fact-checking within email context
- Real-time web search built-in
- Good at drafting factually accurate content quickly
- Pro subscription bundles multiple models (GPT-4o, Claude, etc.)
Weaknesses for email
- Not designed as an email writing tool
- No persistent voice learning
- No Gmail/Outlook integration
- Outputs skew toward informational vs. conversational tone
Pricing: Free tier; Perplexity Pro ~$20/mo
Best for: Research-heavy email replies that need accurate facts or citations
Head-to-head for email
The problem neither solves
Both DeepSeek and Perplexity share the same fundamental limitation for email: they start cold every time. They have no memory of how you actually write: your sentence length, your opener patterns, your sign-off habits, the inside-jokes you use with specific clients. You compensate with elaborate system prompts that you re-paste on every session.
The outputs are good, but they're generically good. Recipients increasingly recognize the cadence of AI-drafted email: the em-dash overuse, the "I hope this finds you well," the verbose sign-off. These tells erode trust in relationship-driven communication.
The alternative is a tool that actually learns your sent email history, not from a one-time prompt, but from the real pattern of how you write. ForthWritedoes this inside Gmail and Outlook directly. You don't tab-switch; you draft in your inbox, and the AI knows your voice because it has read your email history.
Stop patching your prompt. Learn your voice once.
Build a first-person persona prompt that captures how you actually write. Free, in under 5 minutes.