Some time ago I got excited about running multiple Claude coders. This required moving from interactive sessions to claude -p, adding Beads, later a simple Python orchestrator, a UI, an ask_user MCP, and Telegram integration. I can now run coders non-stop, but with intense usage I consume subscription limits within 10 minutes, even when I use workers with the Sonnet model. For some time I have had a few subscriptions, and during the day I could switch subscriptions when one was consumed. But it's still nowhere near enough.
On Monday, Jun 15, Anthropic plans to change the way they calculate limits for claude -p. Rumor on the street is that the tokens we get within the $200 subscription fee are equivalent to $8-12K in API cost. The question is: what do you do when the subscription is not enough for coding?
One approach I am testing right now: I added an opencode coder to my fleet (agent swarm orchestrator app), using a qwen3.6:36B model deployed on a local server with 2x GPU cards with 36GB of memory in total. I talk to the Fable 5 model, describe the goal; Fable does the investigation, slices the problem into small and simple subtasks that qwen can handle, writes detailed subtask descriptions, and creates the required dependencies in Beads. Opencode with qwen3.6 does the job. Fable 5 verifies the work and creates additional tasks if required.
This approach is significantly slower, but I can run it non-stop, overnight, over weekends. I'm also considering testing some cheap model in the cloud.
What is your approach? What do you do when you need more than the subscription provides? What the plan B for case when subscription is gone one day.
0 comments
—
Ask HN: What do you do when you hit Claude subscription limits? · HackerTrans
On Monday, Jun 15, Anthropic plans to change the way they calculate limits for claude -p. Rumor on the street is that the tokens we get within the $200 subscription fee are equivalent to $8-12K in API cost. The question is: what do you do when the subscription is not enough for coding?
One approach I am testing right now: I added an opencode coder to my fleet (agent swarm orchestrator app), using a qwen3.6:36B model deployed on a local server with 2x GPU cards with 36GB of memory in total. I talk to the Fable 5 model, describe the goal; Fable does the investigation, slices the problem into small and simple subtasks that qwen can handle, writes detailed subtask descriptions, and creates the required dependencies in Beads. Opencode with qwen3.6 does the job. Fable 5 verifies the work and creates additional tasks if required.
This approach is significantly slower, but I can run it non-stop, overnight, over weekends. I'm also considering testing some cheap model in the cloud.
What is your approach? What do you do when you need more than the subscription provides? What the plan B for case when subscription is gone one day.