The deepseek harness vs claude code question got a real answer this week, and it is not the one the hype cycle wanted.
We gave both agents the same prompt on the same afternoon.
DeepSeek Harness finished the entire build in 11 minutes.
Claude Code was still working at 30 minutes.
The DeepSeek run cost about 5 cents.
Then we opened both results, and the Claude one was the only one that looked like a real business website.
That is the whole comparison, and everything below is the detail behind it.
The verdict before the detail
Claude Code takes anything a client will see.
DeepSeek Harness takes volume, drafts and anything disposable.
Kasra Dash scored Claude Code a 9 out of 10 after using it daily since launch.
I scored DeepSeek Harness a 7 out of 10, and he thought I was being generous.
Neither of us moved our day-to-day work across.
Both of us moved specific tasks across immediately.
The test we actually ran
One prompt, sent to both, with no context and no follow-up messages.
We asked for a 3D animated accountancy website, plus a simple game.
DeepSeek Harness ran DeepSeek V4 Pro, the model built for it.
Claude Code ran Claude Opus 5 on high.
Frontier model against frontier model, so nobody can say the test was rigged.
Speed: 11 minutes against 30-plus
DeepSeek Harness went from prompt to finished build in 11 minutes.
Not first output. Finished.
Claude Code passed the 30-minute mark and was still going when we moved on to talk about something else.
If you ship for a living, that gap is not a statistic.
It is the difference between five attempts in an hour and one.
Quality: the part that flipped the result
Then we opened the two homepages side by side.
The Claude build had animation that followed the mouse, so the page reacted as you moved across it.
It also referenced things we never typed, including Northwest England, because it carried context from previous work.
That is what an agent that actually knows your business looks like in practice.
The DeepSeek build came out animated and cartoony.
Its Tetris ran perfectly well.
But a cartoon game on an accountancy website is a mismatch, and Kasra called it within seconds.
The Claude game also felt smoother to play.
Be clear about one thing though: neither build was ready to publish.
Both needed more context, more prompting and more rounds before they were worth showing a client.
Old way vs new way
| Old way | New way |
|---|---|
| Pick one coding agent and pay whatever it costs. | Run several engines and route each job to the cheapest one that can do it. |
| Every draft burns premium tokens. | Drafts cost about 5 cents on the cheap engine. |
| One attempt per 30 minutes. | Roughly five attempts per hour at 11 minutes each. |
| The vendor picks your model. | Everything is a plugin, so you pick the brain. |
| Switching tools means rebuilding your workflow. | An orchestrator holds the workflow, engines swap underneath. |
🔥 Want the exact orchestrator setup?
Inside the AI Profit Boardroom I show the full Agent OS build — DeepSeek Harness, Hermes and Claude Code running side by side — plus weekly coaching calls with 4,000+ members.
The token burn nobody mentions
DeepSeek Harness ripped through 483,000 tokens building that one-page site.
Claude was at 48,000 tokens at the 20-minute mark on the same brief.
That is roughly ten times the tokens for the less polished result.
I said on camera that it felt like being cheated, and I stand by that reaction.
The counter-argument is equally real.
DeepSeek is about 56 to 57 times cheaper per token, so the whole build still came in at around 5 cents.
Even burning ten times the tokens, the cheap engine wins the maths comfortably today.
The caveat is that this depends entirely on current pricing, and a very verbose model at a normal price is a much less interesting product.
Head to head
| Metric | DeepSeek Harness | Claude Code |
|---|---|---|
| Time to finished build | 11 minutes | Still running at 30 minutes |
| Model | DeepSeek V4 Pro | Claude Opus 5 (high) |
| Tokens used | 483,000 | 48,000 at 20 minutes |
| Cost of the run | About 5 cents | Roughly 57x more per token |
| Used unprompted context | No | Yes |
| Our score | 7 / 10 | 9 / 10 |
| Maturity | v0.1 developer preview | Mature, shipping |
Why 105,000 people starred it in two days
DeepSeek Harness landed on GitHub and collected 105,000 stars in roughly two days.
That makes it one of the fastest growing open source projects anybody has tracked.
It did not happen because of the build quality.
It happened because the harness is free, open and model-agnostic.
You run it locally, and you decide which brain goes inside it.
If you want the cheapest possible setup you can plug a free model in, and OpenCode works as a free brain inside the harness.
Compare that with a closed product where the model, the price and the roadmap are all decided for you.
Whatever you think of the output, that is a different category of offer.
The v0.1 asterisk
Every criticism in this article carries the same footnote.
This is a developer preview, version 0.1, roughly a tenth of what it will be at a real release.
That is why my score and Kasra’s differ.
He rated what is on screen today, which is fair.
I rated the trajectory, which is also fair.
I also just like what competition does to pricing and to release schedules.
If DeepSeek ships something serious in two or three months, Claude has to answer it, and that is worth more to you than whichever tool is marginally ahead this week.
How I actually use both
I am not picking a side and defending it.
Client-facing work runs on Claude Code, because polish is what gets paid for.
High-volume, repetitive and disposable work runs on DeepSeek Harness, because attempts cost pennies and speed compounds.
The switching is automated rather than manual.
An orchestrator inside my AI agent operating system decides which engine takes each job.
I did not even install the harness by hand — I asked Claude to set it up, test it and wire it in, so I never had to learn another interface.
That is the pattern I would push anybody towards, because it means the next launch is an upgrade rather than a migration.
Three jobs worth moving across today
First drafts of anything. The first attempt is usually wrong, so pay as little as possible for being wrong.
Bulk generation. Page scaffolds, boilerplate, repetitive variations — an 11-minute turnaround at near-zero cost compounds fast here.
Experiments you would otherwise skip. The tests you avoid because the token bill feels wasteful become free.
What should not move is anything a client opens, and our test made that obvious.
🔥 Want my full AI agent training?
The AI Profit Boardroom has the step-by-step builds, 1,000+ AI agents and five coaching calls a week. You can also start free inside the AI Money Lab.
What the harness actually is
A lot of people hear “DeepSeek Harness” and assume it is a model.
It is not.
The harness is the thing that wraps around a model and lets it do work: read your files, run commands, change code, search the web, and keep going without you approving every step.
The model is the brain, and the harness is the body.
That distinction is why this launch mattered more than a normal model release.
DeepSeek shipped V4 Pro and a harness on the same day, so the model finally had somewhere to live.
And because the harness is open source, the brain inside it is your choice rather than the vendor’s.
Why this changes how you budget
Most people budget for AI the way they budget for software, as a monthly line item they try to keep flat.
That framing breaks the moment one of your engines costs a fraction of a cent per attempt.
When a build costs 5 cents, you stop rationing attempts, and the constraint moves from money to judgement.
That is a much better place for the constraint to sit.
The mistake is assuming the cheap engine should therefore do everything.
Our test showed the cheap output landing further from finished, and every extra prompting round costs you something that does not appear on an invoice: your time.
So the rule I use is simple. If the output goes in front of somebody who pays me, it runs on the premium engine. If the output is an input to my own next step, it runs on the cheap one.
That single rule captures most of the saving without any of the risk.
What the test does not tell you
One prompt is one data point, and it is worth saying that plainly.
We ran a single brief, with no context and no follow-ups, on both agents on the same day.
That is a fair snapshot rather than a benchmark suite.
If your work is more repetitive than ours, the cheap engine looks even better than it did here.
If your work leans heavily on accumulated context, the premium engine pulls further ahead than our numbers suggest.
The way to settle it for your own business is to run the same test yourself: one prompt, both agents, no extra context, and watch the token counters as well as the output.
On the DeepSeek side that costs about 5 cents to find out.
There are very few decisions in this industry you can settle that cheaply.
Frequently asked questions
Is DeepSeek Harness better than Claude Code?
Not on output quality today. It is far faster and cheaper, and Claude Code still produced the better website in a like-for-like test.
How long did each agent take?
DeepSeek Harness finished in 11 minutes. Claude Code was still running past 30 minutes on the identical prompt.
How much did the DeepSeek Harness build cost?
About 5 cents, from a $10 top-up, for a 3D animated site plus a working game.
Is DeepSeek Harness free?
The harness is free and open source. You supply the model, and free brains like OpenCode work inside it.
Should I switch from Claude Code to DeepSeek Harness?
Not wholesale. Move volume work across, keep client work on Claude, and let an orchestrator route each job automatically.
About Julian
I’m Julian Goldie, founder of a 7-figure SEO and link building agency (Goldie Agency, 70+ team) and the AI Profit Boardroom.
I have 400K+ YouTube subscribers, 163K X followers and 29K+ Udemy students, and I wrote Link Building Mastery.
I test these tools on real client work before I write about them, which is why this comparison has numbers in it rather than opinions.
Also on our network
The same test, written up from a different angle: juliangoldie.co.uk and goldstarlinks.com.
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
The honest deepseek harness vs claude code verdict is to run both, and let the cheap one do the work nobody sees.
