Claude Code + DeepSeek | 95% cost saving.
Sharing my experimentation, story, and research
Weekly articles on AI, building, and where design is headed. Ready to become AI-native? Join the AI Design Academy →· Connect on LinkedIn →
A lot of my AI Design Academy members have been asking me the same question.
“Is there a cheaper way to use Claude Code?”
They’re learning. They’re vibe coding. They’re building. And they don’t want to burn through credits while figuring things out. That’s a fair concern.
So I went down the rabbit hole this week to find out what actually works. Local LLMs. Gemma, Kimi, DeepSeek. Other cheaper and open source models.
Here’s exactly what I tried, what failed, and what I’d actually recommend.
I started with local LLMs
The first thing I tried was Ollama. It runs models directly on your Mac. No internet needed. Completely free. I tried to pull Kimi first. Turns out the version I wanted needed an Ollama Cloud subscription, so that was out. I switched to Gemma. A 9 GB on-device model from Google. Free. Local.
It worked. Sort of.
My Mac started slowing down right away. The responses were slow. Other apps started lagging. It was barely able to handle simple tasks. I couldn’t even keep working in the background while it was running. I have a decent Mac. It still struggled.
Here’s what I figured out. To actually run local LLMs well, you need serious hardware. Like a Mac Studio Ultra or a workstation with a real GPU. Not a regular MacBook.
A lot of people online hype local LLMs. “Run AI for free on your laptop.” But that’s far from the reality for most of us. Free comes with a cost your time.
Then I looked at cheaper paid options
Anthropic has a few different plans for Claude Code. Claude Pro starts at $20 a month. That’s the entry tier fine for casual use. Claude Max 5x is $100 a month. That’s what I’m on. It gives you 5x the usage of Pro, which is what you need once you’re using Claude Code daily for real work.
Claude Max 20x is $200 a month. That’s the heavy power-user tier 20x Pro usage. Some of the developers I follow are on this one because they’re running Claude Code basically all day.
I’m spending $100 a month on Claude Max 5x. That’s $1,200 a year.
For me, it’s worth it, Claude Code is core to how I build and how I teach. But if you’re a designer trying to learn, a student exploring AI, or a builder still figuring out your stack $1,200 a year is real money. That’s a laptop. And $200/month for the top tier? That’s $2,400 a year.
So I started looking at the cheaper APIs. Kimi. Qwen. DeepSeek.
DeepSeek stood out because I’d already been using it.
A bit of context I run a personal AI agent called Bruno. He’s built on the OpenClaw framework. His job is to follow up on my daily to-dos and keep me accountable. I share what I’m planning for the day, he plans it out, follows up, and holds me to it.
Bruno runs on a separate Mac Mini. He stays out of my way and just runs in the background.
When I first built Bruno, I had him running on the Anthropic API. He worked well, but the cost was adding up fast. Every simple task was using credits. I was building a personal accountability agent, not a production app. The math wasn’t working.
I switched him over to DeepSeek. The cost stopped being a problem. The quality held up. He still works the same way. No drop in performance that I’ve noticed.
That’s when DeepSeek became real for me. Not a benchmark on Twitter. A model I’d actually been using on a real product for weeks.
DeepSeek just dropped V4 and it’s a big deal
Last Friday, DeepSeek released V4. Two versions. Pro for harder work, Flash for fast and cheap. The internet has been talking about it nonstop. Most of the conversation is about geopolitics, benchmarks, or which lab is winning. That’s not the angle that matters to me.
The angle that matters is the price. For context, here’s the rough math in plain language.
Claude has three main individual plans:
Claude Pro at $20/month, about $240/year. Claude Max 5x at $100/month, $1,200/year. (This is me.) Claude Max 20x at $200/month - $2,400/year for the heaviest users.
DeepSeek V4 for the same kind of coding work? Most builders are spending $2 to $10 a month. One developer posted that he ran 412 tool calls in a single day of heavy work and spent under $7 total.
A whole month of casual building on DeepSeek can cost less than a coffee. So we’re talking about the difference between $1,200 a year and roughly $60 a year for similar results on most coding tasks.
The quality is roughly 90 to 95% of Claude on coding tasks. Not identical. Close. For most designers, semi-technical builders, and students learning this stuff that gap doesn’t matter. The savings do.
You can plug DeepSeek into Claude Code itself
Here’s the part that clicked for me.
Claude Code is the best coding interface I’ve used. The shortcuts, the file handling, the agent flow Anthropic nailed the experience. I wasn’t going to give that up.
Turns out I didn’t have to.
Claude Code is more flexible than most people realize. With the right setup, it can run on a much cheaper engine in the background. Same interface. Same workflow. Way lower cost.
That’s the actual play. Not “ditch Claude Code.” Use Claude Code smarter. I wrote a complete guide in AI Design Academy.
My honest take after a week of this
I still use Claude Code with the Anthropic API for my own daily work. I do workshops on Claude Code, so that’s still my setup.
But Bruno ( my Openclaw agent ) runs on DeepSeek. And for my members who want a cheaper way to learn Claude Code DeepSeek + Claude Code is a real option. I’ll share these experiments in the coming day, make sure you subscribe.
What nobody seems to be saying
Most of the takes I’m seeing are either super technical (architecture papers, benchmark debates) or super political . Both miss the point for normal people. For most designers, non-technical builders, and semi-technical people the cheaper models are enough.
We get our work done. We build our websites and apps. We’re not running multi-agent systems 24/7 with edge-case requirements. We just want to ship. The technical crowd hypes the small differences. For the rest of us, the gap doesn’t show up in real work.
The thing I’m actually excited about is the impact on students.
I gave a talk at UC Irvine recently. A lot of the students were worried about cost. They want to learn AI, build with it, experiment with it but the price tag was a barrier. Cheaper models like DeepSeek change that. AI becomes accessible to people who couldn’t afford it before.
Where it might fall short (I’ll keep testing)
I’ve been using DeepSeek heavy on Bruno( Openclaw agent). Haven’t seen a single hallucination yet. The agent is active, performing well, and stable.
But I haven’t stress-tested it on vibe coding yet. That’s the next experiment. So before I claim DeepSeek is bulletproof give me a couple weeks. I’ll post my findings. Subscribe so you don’t miss the follow-up.
What I’m telling my Academy members
I just dropped a complete guide inside the AI Design Academy this week. It’s in the Claude Code track. Members get the full setup. The exact configuration. The model pricing breakdown.
If you’re already a member, it’s waiting for you in the Claude Code track.
If you’re not this is the kind of thing the Academy exists for. Real setups. Real costs. Real builds. I share what’s actually working before the rest of the internet catches up and reduce the noise.
What’s next
Vibe coding with DeepSeek. I want to see if I can actually build full apps using Claude Code on the cheaper model. Real builds. I have workshop coming up on Maven, i’ll share my setup there as well. That’s the experiment I’m running next.
Thank you for reading.
Become an AI-Native Product Designer. Join designers and design engineers inside the AI Design Academy.