Skip to main content
Two limits apply: a free allowance of 3 merged deliveries before you subscribe, and 300,000,000 tokens of model usage per month once you do. Both work the same way when they are reached: newly dispatched work is paused with an explanation on the Issue, and a delivery that is already running is never interrupted.

​
The free allowance

Your first 3 merged deliveries are free. No card, no subscription. A slot is taken when you label an Issue ai-ready, and it is refunded if that delivery does not end in a merge — an Issue that ends ai-blocked, or that you close without merging, gives its slot back. That is why the allowance is described as three merged deliveries rather than three attempts. Counting is per Issue and idempotent: re-labelling the same Issue does not consume another slot. Re-delivering the same Issue after a refund uses the slot it gave back.

​
When the allowance runs out

The next Issue you label ai-ready has that label removed and gets a comment saying the free allowance is used up, with a link to subscribe.
The status page for a tenant whose free allowance is used up

The status page for a tenant whose free allowance is used up

To resume: subscribe, then re-apply the ai-ready label to the Issue. The label is not restored automatically. Deliveries already in flight when the allowance ran out finish normally.

​
The monthly token quota

A subscription includes 300,000,000 tokens per calendar month (UTC). There is no per-token overage billing — when the allowance is used up, new deliveries pause until the next month starts. Usage counts input tokens, output tokens and cache-read tokens. Re-delivering the same Issue accumulates: each delivery is a separate run. As a rough guide, 300M tokens covers somewhere between 85 and 400 merged deliveries depending on how large the tasks are. That range assumes your provider supports prompt caching.

​
Watching your usage

The Monthly usage card on the status page shows tokens used against the limit, how many remain, and a percentage bar. Beside it, Monthly cache hit rate matters more than it looks. Cache-miss tokens cost roughly 50 times more than cache hits, so a rate below 90% burns the quota noticeably faster. If yours is low, it is worth understanding why before assuming the quota is too small. Each work item also has its own usage detail — input, output and cache tokens, runtime, model requests and how many times the run resumed. Work that never reported usage shows —.

​
When the quota runs out

The behaviour matches the free allowance: the ai-ready label is removed from newly dispatched Issues and a comment explains why, naming your actual usage and the cap. In-flight deliveries are untouched. To resume: wait for the next calendar month, or configure your own model key — BYOK usage is measured but never paused. Either way you must re-apply ai-ready yourself.
If Orbi’s usage measurement is unhealthy, the quota check fails closed rather than guessing — new dispatch pauses with a comment saying usage collection is temporarily abnormal. This is deliberate: it will not spend against a number it cannot trust. It clears on its own.

​
What is never interrupted

Neither limit ever stops work in progress. Both guards only act on an Issue that has just been labelled ai-ready and carries no other ai-* label — so a delivery that is ai-in-progress, ai-pr-opened or ai-fix-needed runs to completion regardless. Platform upgrades and deployments have the same property: they never restart a delivery that is already executing.

​
The one-repository limit

Orbi Cloud works in one active repository at a time. You can change which one at any time — see connect a repository — but not while a delivery is in flight.

​
Summary