dax says OpenCode Go broke even two weeks running
The OpenCode founder followed the milestone with public math on how a lab sells a $200 plan without going broke.
dax reports two consecutive break-even weeks for OpenCode Go, and says that matters because everyone else shipping this shape of product is either burning money or silently opting users into a data selling scheme. His words, his numbers, and neither is independently verified. But it is the clearest public accounting anyone in this crowd has given of what it actually costs to resell inference.
we've had two consecutive weeks where OpenCode Go broke even
The $200 plan math
Asked how a lab can sell a $200 subscription to heavy users, dax sketched it out with numbers he labelled as made up. Assume 90 percent margins on Claude. A $200 subscriber who spends $2000 of list-price tokens breaks even. Say the average is $3200, so the lab loses $120 per user. With a million subscribers that is $120M of loss, which $133M of API spend at the same margin covers. His point was not the precision, it was that the loss-leader is smaller than it looks from the outside.
Plenty of people told him he was wrong. He noted a few days later that Anthropic had reported gross margins over 80 percent.
Why cheap tokens should worry you
The warning that travelled furthest was about inference providers underselling the labs. dax says OpenCode has come across sketchy operators with real usage volume, including ones that send fake tool calls to steal information, and ones that sell entire traces on to third parties who resell them again. A single API key sitting in a trace is enough to cause real damage. His rule is simple. If you cannot explain the economics of someone's cheap tokens, do not plug your harness into them.
Underneath all of it is a structural claim worth sitting with. dax argues the only companies able to offer true token based billing are the labs themselves. Everyone else eventually bails out to provisioned capacity because they have to plan around your specific traffic, which means you pay for idle GPUs. So if you want to build on an open source model and scale up and down freely, he says you currently cannot. OpenCode's stated goal is to get large enough to be the first to offer that.
That is also why he is talking about building a GPU cluster. He says every cluster they have looked at is generic, built to support a wide range of workloads rather than specialise, and that is why nobody matches DeepSeek's inference price.
The business read
If your product's unit economics depend on somebody else's subsidised tokens, dax's math is the version of that risk to keep in your head. Break-even at his scale took owning the stack down to the hardware question.
if you cannot explain how someone can offer cheap tokens you should avoid it