Key Takeaways:
💡 Switching to Deepseek’s API can reduce AI implementation costs by up to 90% while maintaining performance comparable to GPT-4.
💡 Deepseek offers global accessibility without VPN requirements and includes built-in reasoning capabilities that streamline complex tasks.
💡 Migration to Deepseek requires minimal code changes due to its OpenAI-like structure, making it a low-risk, high-reward transition for developers.
Introduction

I use chatGPT and Claude API calls for all my automations in n8n and make. I also use API calls for code editing like Cursor. The problem is that it can get EXPENSIVE. USD$0.015 per call might not seem much but when you’re running thousands of calls, it adds up. That’s when Deepseek came out. An A.I. model that rivals gpt-o1. Most importantly, it’s 90% cheaper!
Cost Comparison Summary
Model/Provider | Input Cost (per 1,000 tokens) | Output Cost (per 1,000 tokens) | Notes |
---|---|---|---|
DeepSeek | ~$0.002 | ~$0.002 | Budget-friendly, competitive. |
ChatGPT (GPT-4) | ~$0.03 | ~$0.06 | High performance, higher cost. |
ChatGPT (GPT-3.5) | ~$0.0005 | ~$0.0015 | Cost-effective for basic tasks. |
Claude 3 Opus | ~$0.015 | ~$0.075 | Advanced, higher cost. |
Claude 3 Sonnet | ~$0.003 | ~$0.015 | Balanced performance & cost. |
Claude 3 Haiku | ~$0.00025 | ~$0.00125 | Lightweight, low cost. |
I tested Deepseek with coding tasks and writing. The results were surprisingly good. I actually just started using Thinking Claude to see the A.I. think and Deepseek already has thinking built in.

I spent the whole day changing my n8n, make and cursor to use Deepseek. I’m expecting to see a 90% decrease in my costs. I’m so excited!
The added bonus is that I don’t need to use VPN to access chatGPT and Claude anymore. Users from Asia can now access a world class LLM.
My testing results with Deepseek
I first gave it a coding challenge.
"I have a git tag and I made a mistake. I need to pull it, create a branch from it, update the code, delete the remote and local tags and then push it. Can you give me the commands needed?"
4o , O1 and Claude both failed because they forget I must have a local tag and/or a remote tag to check it out, so the commands they give me simply ignore this and fail because the tag exists. On a second run, O1 did give me all the steps, so it’s half a win.Online DeepSeekR1 got it right first try.
Noticeable decrease in performance with Claude 3.5 Sonnet and GPT-4o
I’ve been noticing that Claude and GPT are showing defensive behaviors like a human. When I ask it to correct something, it’s starting to make up answers instead of trying to solve it properly. One way of solving it is to add a “thinking” process. Thinking Claude does this and so does Deepseek.
So far, both prompting and coding scenarios have performed well for me.
How to get Deepseek if you’re banned

Kevon left a comment on my threads account saying he can’t get Deepseek API access. If you’re unable to get access to Deepseek API. I recommend using Openrouter. It also allows you to change models in your n8n and make quickly.

My thoughts on Deepseek
Deepseek is a game changer. It’s free. It doesn’t require VPN access. It’s open source.

By making A.I. API calls cheap, it’s going to unleash MORE people to use A.I.. Every A.I. company is now looking at Deepseek’s code and copying some of it. The price of API calls will drop for everyone. This is good for us! Make use of it.