Migrating from OpenAI to DeepSeek: A Practical Guide

Last updated

DeepSeek exposes an OpenAI-compatible API, so moving from OpenAI is one of the easiest migrations there is: point the OpenAI SDK at DeepSeek’s base URL, swap the key, and change the model. No request/response remapping.

API concept mapping

OpenAI → DeepSeek (as of June 2026).
ConceptOpenAIDeepSeek
SDK openaiopenai (same) with base_url
Base URL defaulthttps://api.deepseek.com
Auth OPENAI_API_KEYDEEPSEEK_API_KEY
Endpoint /chat/completions/chat/completions (same)
Model gpt-5.4 / minideepseek-chat (non-thinking), deepseek-reasoner (thinking)
Response choices[0].message.contentchoices[0].message.content (same)
Tools / streaming standardstandard (compatible)
The whole migration

Keep the OpenAI SDK. Set base_url="https://api.deepseek.com", use your DeepSeek key, and set the model to deepseek-chat or deepseek-reasoner. Done.

Before / after

Before — OpenAI 
from openai import OpenAI
client = OpenAI()  # OPENAI_API_KEY

resp = client.chat.completions.create(
  model="gpt-5.4",
  messages=[
      {"role": "system", "content": "You are concise."},
      {"role": "user", "content": "Explain rate limits in one line."},
  ],
)
print(resp.choices[0].message.content)
After — DeepSeek (same SDK) 
import os
from openai import OpenAI

client = OpenAI(
  api_key=os.environ["DEEPSEEK_API_KEY"],
  base_url="https://api.deepseek.com",
)

resp = client.chat.completions.create(
  model="deepseek-chat",                 # or "deepseek-reasoner" for thinking
  messages=[
      {"role": "system", "content": "You are concise."},
      {"role": "user", "content": "Explain rate limits in one line."},
  ],
)
print(resp.choices[0].message.content)

Gotchas

Watch for these
  • Model names drift. DeepSeek has renamed model ids over time (e.g. the deepseek-chat / deepseek-reasoner aliases map to the current V4 generation). Confirm the live id in DeepSeek’s docs before shipping.
  • reasoner returns reasoning content. The thinking model exposes a separate reasoning field and longer outputs — budget tokens and don’t feed reasoning back as input.
  • Re-tune prompts and re-run evals. Compatible API ≠ identical behavior; outputs differ from GPT models.
  • Some OpenAI-only params (certain response_format / tool edge cases) may differ — test tool calling specifically.

What to do next

  1. Centralize your model id and base URL in config so the swap is one line.
  2. Shadow-run DeepSeek on a slice of traffic and diff outputs.
  3. Price it first — DeepSeek is often the cheapest option; confirm with the cost calculator.
  4. Going back later? See DeepSeek → OpenAI.

Frequently asked questions

Is DeepSeek really OpenAI-compatible?
Yes — it implements the OpenAI chat-completions API, so the official OpenAI SDKs work by setting base_url to DeepSeek and using a DeepSeek key. Most apps migrate with config changes only.
Which DeepSeek model replaces GPT-5.4?
Use deepseek-chat for standard (non-thinking) responses and deepseek-reasoner for chain-of-thought tasks. Verify the current model ids in DeepSeek's docs, as they're periodically renamed.
Will my tool calling and streaming keep working?
Generally yes — both are supported through the compatible API. Still test tool/function calling end to end, since edge-case parameter handling can differ between providers.