if you want AI agents writing code for you around the clock, you’re choosing between two billing models.

one runs the agents on someone else’s cloud and meters what they burn: a quota per plan, pay-as-you-go above it. the other runs them on a subscription you already pay for, on a box that’s yours.

the feature lists blur together. this is the split that actually shows up on your bill and in who controls the machine.

the two models

the metered-cloud model. you hand a coding task to a hosted service. it plans, writes, tests, and hands back a PR, all on its infrastructure. you get a usage quota, and past it you pay as you go. nothing to run yourself, and the meter tracks every token the agents spend. clean if you’d rather never think about servers or subscriptions.

the own-subscription model. you get a server of your own and run your AI agents on it as a team. they live on the box around the clock, message each other, and reach you over telegram. the agents run the official Claude CLI on your own Anthropic Pro or Max plan, so the work they do is covered by the subscription you’re already paying for. no second meter ticking on top.

5dive is the second model. the rest of this is the honest case for when each one wins.

price

	own-subscription (5dive)	metered cloud
starting price	$19/mo flat	~$20/mo + usage
billing model	flat, your own VM	plan quota, pay-as-you-go above it
AI inference cost	your own Claude subscription	metered per use, billed by the provider
heavy usage day	costs the same	costs more
who runs the compute	you (dedicated VM)	the provider’s cloud

the starting numbers land close. the shape of the bill doesn’t.

a metered plan bills the model layer for you. that’s genuinely convenient: no keys, no subscription to manage, one invoice. the tradeoff is that a heavy day, an agent thinking and iterating for hours, spends more. the own-subscription model puts the AI cost on the Claude plan you already hold, so an agent working all day costs the same flat rate as one working an hour. no per-task quota, no meter.

the four real differences

1. your subscription vs metered usage

this is the big one. 5dive runs the official Claude CLI as a live interactive session on your own Pro or Max plan. what your agents do all day rides on the subscription you already pay for.

a metered service runs inference on its own cloud and bills it back to you, quota plus overage. it’s a fine model, it’s just a different one. with a subscription you’ve capped the cost; with a meter the cost moves with the work.

2. a box you own vs someone else’s cloud

5dive runs on a per-user VM that’s yours. you can ssh in, read every log, inspect state, install whatever you want. a hosted service runs the agents on its own managed cloud. that’s less for you to babysit, which plenty of people want. it’s also not a machine you control, and your keys and work live in someone else’s dashboard.

3. a live agent vs a task you hand off

most hosted agents are async: you kick off a task, it runs on their servers, it pings you when the PR is ready. clean for “go build this, tell me when it’s done.”

5dive lives on telegram as a live agent. you message it, it replies in real time, runs commands, shows you what it’s looking at, takes a follow-up mid-task. it’s a conversation with something that’s already running.

4. a team across roles vs a single-purpose coder

most coding services do one kind of work: engineering. every use case is a coding task.

5dive runs agents as distinct, named teammates with different jobs: a researcher, a marketer, an ops bot, an engineer, each its own identity, each able to call the others through the CLI itself. you give them an org chart and a shared task queue, and they hand work down it, with human approval gates where you want them. it’s a company. the coding agent is one seat at the table.

one more thing: it’s open source

5dive’s host toolchain is MIT and on github. you can read it, run it on your own hardware, fork it. most hosted services are closed. for a lot of buyers that doesn’t matter. for the ones it does, it matters a lot.

who each model is for

pick the metered cloud if you want to hand off a coding task and get a tested PR back, you’d rather not run any infrastructure or think about subscriptions, and paying per use fits how you work.

pick the own-subscription model if you want to own the box, keep AI costs on the Claude plan you already have, run a team of agents that do more than code, and reach them as a live presence over telegram. and if you’d like the option to read or self-host the whole thing.

both are real ways to run agents. they’re just different bets on who owns the machine, and who pays the meter.

---

5dive runs your AI agents on a server you own, on your own Claude subscription, reachable from telegram. from $19/mo. start here, or read the source at github.com/5dive-ai/5dive.