I said out loud to myself, "Claude Cowork is great, but I'm paying for it." Then I killed it. Well, not killed — I replaced it with DeepSeek running through OpenClaw, and it turned out to be one of the more sensible cost decisions I've made this year.
Here's the honest version: it's a little less smart than Opus 5. Not far behind — but on the gnarly edge cases, I notice it. The trade is worth it. This is the story of that trade, tested on real tasks, not a benchmark.
The setup
I was running Claude Cowork as my main agent alongside a bunch of other coding agents. It was good — Opus 5 is genuinely impressive. But it has a cost that adds up when you run it every day, across multiple projects, and especially when you parallelize.
I switched the agent layer to OpenClaw with DeepSeek v4 flash as the model. The key move: opencode go serves DeepSeek v4 flash free up to a point — which covers most of my day-to-day agent work. When I need a guaranteed paid path, DeepSeek via OpenRouter is cheap anyway (about $0.14 in / $0.28 out per 1M tokens).
So the stack is: OpenClaw as the agent harness, DeepSeek v4 flash as the brain, opencode go as the free model lane.
What I tested it on
I didn't just run hello-world prompts. I ran a week of real work through it. Four tasks that actually exercise an agent:
Chrome connection / browser automation. This was the one I was most worried about. Agents that can drive a browser can break a lot of things if they're sloppy. I had it connect to a live Chrome instance and work through real pages — login flows, form filling, screenshots, verifying rendered output. It handled it. Not flawless, but it reasoned through the steps and corrected itself.
A Laravel web app (cijenehr, with LearnCoachAssist in the mix). Full-stack work: Blade templates, scraping adapters, deploys, fighting framework gotchas. DeepSeek is solid at Laravel — it knows the framework well enough to be productive, and it caught real bugs (including some genuinely cursed Blade comment issues).
A scripted media pipeline (agicad). Automated short-form video generation publishing to Instagram, YouTube, and Facebook on a cron. Debugging a pipeline that "just stopped posting" — turned out the script library was dry, not a code bug. It reasoned to the right diagnosis instead of rewriting code it didn't need to.
A multi-agent server setup. Coordinating OpenClaw + a Python agent (Hermes) on a remote server, migrating models, fixing auth failures, wiring fallbacks. This is where the "a bit less smart" showed up — multi-step infrastructure reasoning is exactly the kind of edge case where a stronger model pulls ahead. But it got there, and the cost difference was enormous.
Where it's weaker
Let me be straight, because this is the part that matters.
DeepSeek v4 flash is not Opus 5. On straightforward, well-scoped tasks — writing code, fixing bugs, browser work — it's maybe 90-95% as good. The gap shows on:
- Long, multi-step infrastructure reasoning where you have to hold a lot of state in your head
- Tasteful judgment calls on ambiguous architecture decisions
- The last 5% of edge-case handling that a top-tier model just instinctually nails
But here's the thing: for most real work, "90% as good at 0% of the cost" is a trade I'll take every single time. And when I absolutely need the sharper model, I can reach for it — I just don't need to keep it running for everything.
The cost breakdown
Let me be concrete about why this matters:
- opencode go — DeepSeek v4 flash free up to a threshold. This covers the bulk of my daily agent work.
- cline pass — a solid flat $5/month plan if you want a predictable subscription instead of worrying about free-tier limits.
- DeepSeek via OpenRouter — ~$0.14 in / $0.28 out per 1M tokens as the paid floor.
Compare that to a premium subscription running every day. The savings are basically the entire point. I'm running multiple agents in parallel — that's where the cost compounds, and where going cheap pays off the most.
Try it
If you're running Claude Cowork or another premium agent and thinking "this is great but it's costing me," switching to DeepSeek + OpenClaw is a low-risk experiment. You keep the same agent workflow, you just swap the brain for something cheaper.
Getting started with opencode go is free — you can try it before spending anything:
If you go through that link, I get a small referral credit — it helps me keep writing these posts. No pressure, and honestly the free tier is enough to decide if it works for you.
It's a bit less smart than Opus 5. It's not far behind. And it works. That's a trade I'm happy to make.
Top comments (0)