Malte Ubl ran 24 DeepSeek sub agents under Opus 5.5 for $9
A frontier model doing the orchestration, cheap open weights doing what he says is 99 percent of the work.
Malte Ubl says he spent about $9 on DeepSeek while running 24 sub agents at once, each in its own Vercel Sandbox so they stay off his laptop. The setup is deepseek-v4.1-flash-1017 doing the sub agent work inside Claude Code, with Opus 5.5 sitting on top as the orchestrator. His write-up puts his previous spend at about $2K per day in list-price tokens.
until now I have spent like $9 on deepseek which does 99% of the work
The one practical gotcha he flags is a Claude Code detail. The sub agent tool does not take a tool argument, so he asked the model to write a `.claude/agents/` file to define the sub agent instead.
The shape matters more than the number
One person's day of work is not a benchmark, and Ubl does not say what the sub agents were building or how the output held up. What is worth copying is the structure. An expensive model that is good at planning and delegation sits in the orchestrator seat, and the bulk token spend, the file reading, the edits, the grunt work, goes to a cheap open-weights model that does not need to hold the whole plan in its head. Ubl's claim is that DeepSeek did 99 percent of the work in that arrangement, and that it worked wonderfully with Opus 5.5 coordinating.
That is a different trade from the usual one. Most people choosing a cheaper model swap it in everywhere and accept worse reasoning across the board. Splitting by role means the quality drop only lands on tasks that are already narrow and already bounded by the orchestrator's instructions.
The 24 parallel sandboxes are the other half of it. At $2K a day in list-price tokens, running two dozen agents at once is a budget decision as much as a workflow one. At $9, the constraint moves to how many sandboxes you want to pay for and how much review you can absorb.
DeepSeek V4.1 Flash has been in the conversation for speed as well as price, with Fireworks earlier claiming it served the model at 1,521 tokens per second. Fast and cheap is a reasonable profile for a sub agent that mostly needs to execute a clear instruction and report back.
Ubl's closing line is aimed at people already on a flat-rate plan. Even with a Max subscription, he writes, imagine how much more you could do. Nobody has posted a quality comparison against running Opus end to end, so the honest read is that one engineer found the split worked for a day of his own work and the bill was two orders of magnitude smaller.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.