Switching from ChatGPT to DeepSeek: What Changes, What Doesn't

Whether you're switching by choice or because OpenAI just cut you off, moving from ChatGPT to DeepSeek is surprisingly smooth. The API format is compatible, and for reasoning and chat tasks DeepSeek holds up well — though it's not the strongest option for coding specifically.

Here's what actually changes and what doesn't.

What Doesn't Change

API format

DeepSeek uses an OpenAI-compatible API. If your application calls the OpenAI API, switching to DeepSeek typically means changing two things: the base URL and the API key. Your existing code, prompts, and integrations stay the same.

Libraries like the OpenAI Python SDK work directly with DeepSeek by setting a custom base URL. Most wrapper libraries and tools (LangChain, LlamaIndex, Cursor, etc.) have a "custom endpoint" option that works identically.

Prompt structure

Chat completion format — system message, user message, assistant message — works the same way. Your existing prompts will work without modification in most cases.

Streaming

DeepSeek supports server-sent events streaming, same as OpenAI. Real-time token streaming works identically.

Function calling / tool use

DeepSeek supports function calling with the same format as OpenAI. Define your tools in the request, get structured function call responses back.

What Changes

Model selection

Instead of gpt-4, gpt-4-turbo, or gpt-3.5-turbo, you're choosing between DeepSeek-V3.2, DeepSeek-R1, DeepSeek-Coder, etc. The naming is different but the concept is the same: pick the model that matches your task.

Rough mapping:

• -GPT-4 / GPT-4o → DeepSeek-V3.2 (general purpose)
• -GPT-4 for reasoning → DeepSeek-R1 (chain-of-thought reasoning)
• -GPT-3.5 Turbo → DeepSeek-V3 or smaller variants (cost-optimized)

Pricing

DeepSeek is significantly cheaper — V3 at $0.14/M input tokens vs GPT-4's ~$1-3/M. (Source: DeepSeek Pricing) If cost was a concern with OpenAI, it's much less of one with DeepSeek.

Response style

DeepSeek's responses tend to be slightly different in style. They're generally more concise and less "chatty" than ChatGPT. For coding tasks, many developers prefer DeepSeek's more direct style. For creative writing, you might notice the difference.

Chinese language capability

If any part of your application touches Chinese text, DeepSeek is stronger than GPT-4. This matters a lot in Hong Kong where applications frequently handle both English and Chinese.

Reasoning approach

DeepSeek-R1 uses explicit chain-of-thought reasoning that's visible in the response. This is different from GPT-4's implicit reasoning. For complex problems, R1's step-by-step approach often produces better results, but responses are longer.

Migration Checklist

1. Get your API key from platform.deepseek.com

2. Update base URL in your application configuration

3. Update model names from gpt-4 to your chosen DeepSeek model

4. Test your prompts. Most work as-is. If you see quality differences, minor prompt adjustments usually fix them.

5. Update error handling. Error codes and rate limit responses may differ slightly. Test your retry logic.

6. Benchmark. Run your test suite against DeepSeek and compare results. For most applications, quality is equivalent or better. For edge cases, you may need to adjust.

What's Actually Better

Some things are genuinely better with DeepSeek:

Reasoning. DeepSeek-R1 with explicit chain-of-thought reasoning is strong on math and multi-step problem solving. The step-by-step approach catches errors that GPT-4's implicit reasoning misses.

Chat/general use. For conversational AI and general-purpose tasks, DeepSeek is solid and the experience that made it famous.

Cost. 5-10x cheaper per token. At production scale, this is significant.

Openness. MIT license means you can self-host, fine-tune, and inspect the model. This isn't possible with GPT-4.

No geographic restrictions. It just works. No VPN, no workarounds, no anxiety about your API access being revoked.

What's Not as Good

Being honest:

Multimodal. GPT-4o's image understanding is strong. DeepSeek-V3.2 supports tool-use natively but GPT-4o still has an edge in some image tasks.

Ecosystem. OpenAI's ecosystem (plugins, GPTs, etc.) doesn't exist for DeepSeek. You're working with a straightforward API, not a platform.

Documentation. OpenAI's documentation is more extensive. DeepSeek's docs are adequate but less polished.

Coding. For coding-heavy work, DeepSeek can feel slow and isn't the top performer — models like MiniMax M2.5, GLM-5, and Kimi K2.5 score higher on SWE-bench. If coding is your primary use case, consider those alternatives or Cursor's Auto mode.

For general-purpose and reasoning tasks, the switch is straightforward and the cost savings are real.

Sources

• -DeepSeek API Documentation
• -DeepSeek API Pricing
• -DeepSeek-R1 — Hugging Face
• -DeepSeek-R1 MIT License — GitHub
• -DeepSeek V3 vs R1 Comparison — PromptLayer
• -Complete Guide to DeepSeek Models — BentoML

Made the switch from ChatGPT to DeepSeek? We'd love to hear how it went. Subscribe to the Hong Kong AI Podcast or reach out at contact@hongkongaipodcast.com.

Get notified when we publish new articles and episodes. No spam, just signal.