Skip to contentdocs.Growther.si
Download C5

Budgets & costs

Ops is where you keep an eye on the money. It answers one question: what is this costing me, and where is it going?

Why budgets matter #

Agents use SI models, and most models cost money per use. An agent working on a big job can make many calls. Without a limit, a runaway task could spend more than you meant to.

Budgets stop that. You set a ceiling, and C5 will not go past it.

Setting a budget #

  1. Go to Settings › Budget.
  2. In the Hard Budget Cap panel, set the amount and the period, and whether reaching it notifies you (Notify) or blocks new work (Block).
  3. Save.

Ops › Token Budget then shows what is being used and spent against it. Ops itself only reports: the limits live in Settings.

You can set limits for the whole system, or for one project.

How the cap counts #

The cap adds up everything spent since the current period began, by this machine's own calendar: a daily cap starts again at local midnight, a weekly or monthly one on the day you chose, wherever in the world the machine is. Every day's spending counts, the classifier's questions to a model your provider bills included, and Ops › Token Budget shows the same days: today's use, the last 7 and 30 days, and the month. So do the spending figures on the provider cards in Settings › Integrations and the token costs in the ROI reports. A swarm agent's spending is filed under the local day too; what earlier versions filed for it under the UTC day stays there, so around midnight a day can hold a little of its neighbour's.

Note Up to v329 the cap counted the wrong days in two ways. The installed app filed each day's use under a date written the American way (10/22/2026), which the cap could not put in order: it left out most of that spending, could count some older days in every period, and Ops › Token Budget showed nothing for today. And in a time zone ahead of UTC — most of Europe, Asia and Australia — every period began a day early, so a daily cap also counted yesterday's spending. Both are fixed. On the first start after updating, C5 refiles the use earlier versions recorded that way under the right day, so it shows again and counts in the right period.

The voice budget is separate, and lives elsewhere #

Speech has its own limit, in the Voice Spending Limit panel under Settings → Budget — not on this page. It is separate on purpose: voice is charged by the minute of audio and by the character of speech read aloud, which is a different shape of spending from the work your agents do, and most people want a much smaller ceiling on it.

It works the same way as the main cap — a switch, an amount, and a choice of daily, weekly, monthly or yearly, resetting on the day you pick. C5 works out what a request will cost before making it and refuses rather than overshooting.

Two things never count against it:

  • The built-in on-device engine is free, so it keeps working even when the voice budget is spent. If you want dictation that cannot be interrupted by a limit, put that provider first.
  • Silence is never sent. A recording with nothing in it is discarded in the browser and costs nothing.

See Voice for the providers themselves.

What happens at the limit #

C5 does not just stop dead. As you get close, it warns you. At the limit, it pauses new work and tells you clearly what happened.

Nothing is lost. Paused work waits. Raise the limit or wait for the next period, and it picks up again.

Tip Set your first budget low — lower than you think you need. It is easy to raise later, and a low ceiling teaches you what things actually cost.

Seeing where money went #

The Ops charts break spending down by:

  • Time — today, this week, this month
  • Task — which jobs cost the most
  • Model — which SI models you are paying for

If a number looks wrong, click into it. You can follow the cost all the way down to the task that caused it.

Other things in Ops #

Tab What it is for
Token Budget Daily and weekly token use, spend trends, and a breakdown by coordinator.
Triage Queue health at a glance: pending jobs, background jobs waiting for a gap in real work (and how many expired unrun after 7 days), saga rollbacks, how effective triage briefs are, and a forecast of budget impact.
Memory What C5 has remembered from your past work.
Audit A record of what happened and when, including what C5 did about stuck runs. See Audit trail.
Learning Reward signals by agent and gate, so you can see which prompts and routes are improving.
Classifier How the classifier is doing: its decisions, its speed, the calls it replaced and the labels waiting. Whose answers switches between the built-in model's record and a chosen model's. Only accounts that may view Settings see it.
Errors Problems that came up, grouped so you can spot patterns. Cluster Merge Candidates lists groups that look like the same error; merging one into another, or dismissing the pair, needs Edit on Sidebar: HITL Button & Modal in Settings › Access, and anyone else sees the list without those buttons.

Background LLM Jobs, on Triage, counts the background improvement work waiting for a gap in real work — delegation autopsies, skill crystallisation, sparring matches — and how many expired: a job still waiting after 7 days expires without running. It is read when you open Ops.

Self-improvement milestones #

As your agents get real work done, C5 passes milestones. Most of them unlock an advanced self-improvement feature, held back until there is enough history to be helpful rather than guessing early and spending your budget on it. One of them simply marks the point where something you have been doing has built up enough evidence to be worth reading.

When you reach a milestone, a star badge appears in the bottom left of your sidebar. Each badge is independent — you will see whichever ones you have earned, in any order.

Milestone Unlocks at Sidebar badge What it means
Recipes In Use 25 recipe uses Gold star Recipes have planned 25 tasks. Each one now carries a measured verdict on whether using it beat planning from scratch.
Retrospective Loop 50 completed tasks Violet star Analyzes failed tasks to suggest new review rules and compare rule similarity. You review them in Human review (HITL); one nobody reviews for 30 days expires.
Reward Model Promotion 100 completed tasks Blue star Uses historical data to automatically promote experimental prompt variants when they beat the baseline.

Note that the three count different things. The Retrospective Loop and Reward Model milestones count completed tasks, so they arrive in that order as you work. Recipes In Use counts times a recipe was applied to a task — something you start doing only once you have recipes to apply — so it can arrive before either of the others, or long after. Once earned, a badge stays earned.

Tracking your progress #

The two task-count milestones show their progress in Settings › Budget:

  1. Go to Settings.
  2. Open the Budget tab.
  3. Look at the Retrospective Loop and Reward Model Promotion cards to see your current completed task count (e.g. 34/50 or 82/100).

For Recipes In Use, the count you are watching is in Analytics › Recipes, under Blueprint Invocations — the total number of times a recipe has been applied to a task.

Clicking a star tells you what it unlocked. The Retrospective Loop and Reward Model stars take you to Settings › Budget, where you can turn either feature off to manage token usage. The Retrospective Loop's switch there needs Edit on Sidebar: HITL Button & Modal in Settings › Access. The Recipes In Use star takes you to Labs › Recipes instead, because it does not switch anything on — it marks the point where there is finally something to look at.

Using recipes #

Recipes are offered, never imposed. When you create a task — in Chat, or in the task editor on the board — C5 compares what you have written against the tasks each recipe was distilled from, and if one clearly fits it appears as a suggestion beside the recipe picker. Taking it is up to you; ignoring it costs nothing.

Every recipe you apply is then measured. C5 compares the quality of the runs that recipe planned against what tasks of the same complexity score when planned from scratch, and shows the result on both Labs › Recipes and Analytics › Recipes as a figure like +12.4 vs from scratch. A recipe with no measurement yet shows nothing at all rather than a zero — no reuses means no verdict, which is not the same as a bad one. See Labs for where recipes come from.

Keeping costs down #

A few habits help:

  • Use smaller models for simple work. Not every task needs the most expensive option. See Model providers.
  • Be specific. A clear request takes fewer tries than a vague one.
  • Check your schedules. A frequent schedule is usually the biggest line on the bill.
  • Run models locally for routine work, and save paid models for the hard parts.
  • Know what runs in the background. C5 improves itself in the gaps between real work — autopsies of delegated work, skill crystallisation, sparring matches, alert triage — and that work uses tokens too, on the models in the System Priority column of Settings › Integrations › Coordinator Priority. It waits while chats, tasks or schedules hold the models, and the prompt optimizer waits as well. At a global cost cap set to block, background jobs stay queued, and expire after 7 days.

Where to go next #

  • Labs — recipes, skills, and prompt variants.
  • Analytics — track agent performance over time.
  • Settings — full configuration map.
Edit this page on GitHub ↗Last updated Nov 20, 2026