Cost, Budgets & Credits

An agent run costs you in two separate ways, and Fruxon tracks both:

The two never mix. Spend limits count model cost only, and the credits a step draws don't depend on which model you use or how many tokens it reads.

Model cost

How cost is calculated

Every time an agent calls a model, Fruxon records the token usage and prices it with that model's rates. A run's trace breaks the cost down like this:

Tool cost is the cost of media tools that bill by usage rather than by tokens, such as audio transcription and image analysis. It's included in the agent's total cost, and the Tool usage card on the agent's Overview breaks it out.

Pricing tiers

Some models have special pricing that Fruxon applies automatically:

• Long-context pricing: when a request goes over the model's long-context threshold (for example, 200K tokens), the provider's higher long-context rates apply.
• Prompt caching: tokens served from the provider's cache are charged at the cached-input rate instead of the standard input rate.

Where you see cost

• One run: the Executions list shows each run's cost. Open a run to see its trace with the breakdown above.
• One agent over time: the Cost card on the agent's Overview shows the total for the selected time range, the change from the previous period, a chart, and the run and token counts. Its Production / Dev switch shows test spend instead of live traffic; only editors can switch to Dev. The card also carries the month-end forecast.
• By model: the Model usage card on the agent's Overview.
• By tool: the Tool usage card on the agent's Overview.
• By environment: the agent's Environments view. See Observability by environment.

The cost API can also scope these figures to a single agent revision.

Custom pricing

If you've negotiated rates with a model provider, override the built-in pricing so cost reports and spend limits reflect what you actually pay.

• For the whole workspace: Settings → Model pricing lists every model with its rates, marked Default or Override. Click Edit pricing on a row to set your own rates. To schedule a change, set Effective from. Reset to default removes the override.
• For one agent: set an agent-level override through the API. It takes precedence over the workspace override for that agent, and needs the Admin role on the agent's Application.

For models without built-in pricing, such as a custom Hugging Face endpoint, add a workspace or agent override through the API. If no override applies, Fruxon uses its built-in pricing, which follows the providers' published rates.

Built-in pricing

Fruxon has built-in pricing for models from:

• OpenAI: GPT-4o, GPT-4.1, GPT-5.x and GPT-6 models, and OpenAI embeddings
• Anthropic: the Claude Fable, Opus, Sonnet, and Haiku families
• Google Gemini: Gemini Pro and Flash, and Gemini embeddings
• Google Vertex AI: Claude on Vertex, Google embeddings, and hosted open-weight models such as DeepSeek, Qwen, gpt-oss, and Kimi
• xAI: the Grok series
• AWS Bedrock: Anthropic Claude, Amazon Nova, and gpt-oss
• Microsoft Foundry: OpenAI and Claude models, at OpenAI's and Anthropic's list rates (what Foundry charges for Global Standard deployments; Data Zone and Regional deployments cost more)
• DeepSeek: DeepSeek V4
• VoyageAI: embedding models

Spend limits

Spend limits (budgets, in the API) cap an agent's model cost. They're set per agent, in USD only, and checked before each run starts.

Production and Dev

Every agent has two independent sets of limits, one for each kind of run:

Test runs can't use up the Production limits, and a tight Dev limit never blocks live traffic. Each set is shared by everyone who works on the agent, not kept per person.

Default limits

Every new agent starts with three agent-wide limits on both Production and Dev. Each one blocks runs when it's reached and sends alerts at 80% and 100%:

You can't remove these three limits. You can raise or lower them, or set one to Notify only so it alerts without blocking. The windows have to stay in order: the daily limit can't be higher than the weekly one, and the weekly limit can't be higher than the monthly one.

Windows

Windows follow the calendar in UTC. They don't roll:

• Daily: midnight to midnight UTC.
• Weekly: Monday 00:00 UTC to the following Monday.
• Monthly: from 00:00 UTC on the 1st of the month.

Spend for a window starts again from zero when the next window begins. All of an agent's limits apply at once, so a run has to fit under every one of them, and the tightest limit trips first.

Change a limit

Open the Spend limits drawer from any of these places:

• The spend chip in the agent's header. It shows the tightest limit, for example DAILY $1.20 / $5.00, or Limits if none is set. Hover it to see all three windows.
• The month strip on the Cost card (see Month-end forecast).
• The budget banner on the agent's Overview.

At the top of the drawer, choose Production or Dev. Each window shows what's been spent against its limit, whether it Blocks or Notifies, and when it resets. Click Edit on a window:

Then click Save limit. If a window shows No daily limit set (or weekly or monthly), click Set limit to add one.

Changing a limit needs the Admin role on the agent's Application. Changes are recorded in the Audit Log.

When a limit is reached

Fruxon checks the limits before each run starts. When spend in a blocking window has reached its limit, the next run is refused before it starts, and it costs nothing. Because the check happens before a run, the run that crosses the limit still finishes, so spend can end up slightly above the limit.

An API caller gets HTTP 429 with the code AGENT_BUDGET_EXCEEDED and a message that names the window:

This agent has reached its daily cost budget limit ($5.12 of $5.00 spent). Contact the agent admin to increase the budget, disable enforcement, or wait until tomorrow (budgets reset at midnight UTC).

If more than one window is over its limit, the message names the shortest one, because it resets first. Elsewhere in Fruxon:

• Refused runs appear in Executions with the status Blocked, or Blocked ×12 when repeated attempts are grouped. Hover the status to see the reason.
• The agent's Overview shows a banner, Runs paused — daily spend limit reached, with an Adjust limit button. Once a blocking limit reaches 90%, the banner reads Approaching the daily spend limit instead, with Review limits.
• An automation stage that runs the agent pauses until the window resets. See Spend limits and paused stages.

Runs are accepted again as soon as the window resets, you raise the limit, or you switch it to Notify only.

Alerts

When spend crosses an Alert at percentage, Fruxon emails the Admins of the agent's Application who have Budget Alerts turned on. It's on by default. The email names the agent, the window, and the environment, for example Budget Alert: Support agent has reached 80% of its daily budget.

Each percentage sends at most one email per window. If a single run pushes spend past several percentages at once, you get one email, for the highest.

Limits for one environment

When one agent serves several customers through environments, you can also cap a single environment's spend. Set it through the …/agents/{agent}/budget API with an environment query parameter. Use origin for the bucket and period in the body for the window.

A run in that environment has to fit under both its environment's limits and the agent-wide limits, and its spend counts toward both. Environment limits are optional, so you can delete them. They don't appear in the Spend limits drawer, and their alert emails name the environment.

What counts toward a limit

• The model cost of every run in that bucket, priced at your rates, including custom pricing.
• Media tool cost.
• The model cost of scoring a run with a quality check, which counts toward the same bucket as the run.

Run credits never count toward a spend limit. A collection pipeline's Per-run budget cap is a separate limit: it halts one pipeline run, while each item run still counts against the agent's spend limits.

Month-end forecast

The Cost card's header shows this month's spend against the monthly limit, for example $12.40 / $50.00, for the bucket the card is showing. A dot next to it turns amber when the month is heading for the limit and red once it's over. Click it to open the month strip:

• This month shows the spend so far and where the month is heading, for example $12.40 spent · on pace for ~$31 by Nov 1. The projection uses the agent's spending pace over the last 7 full days. Hover it to see the range between a slow and a busy pace.
• The bar shows what's been spent, the forecast range, the likely month-end figure, and the monthly limit as a dashed line.
• A note appears when a busy pace could reach the limit before the month ends, for example A busy stretch could reach the budget around Oct 24, or once the month is over its limit.
• The button on the right, for example $50.00 budget, or Set a budget when no monthly limit exists, opens the Spend limits drawer for the bucket the card is showing. To change the Dev limits from here, switch the card to Dev first.

The month is the calendar month in UTC, the same as the monthly limit. The forecast appears once the agent has at least five full days of spend to read a pace from.

Run credits

Run credits are how Fruxon charges for its own work. Each billable step a run takes draws credits from your workspace's balance. Credits are separate from model cost: when an agent uses your own provider keys, the provider bills you for tokens, and credits pay for Fruxon.

What a step costs

Everything else is free, including:

• Any step that errors, of any type, and every retry after it.
• A quality check that just records a score.
• Time spent waiting for a human approval.
• Context summarization.
• Receiving the message and sending the reply back on your channel.

A few more rules:

• The price is fixed when the step runs. The trace keeps the credits each step was charged, so a later price change doesn't rewrite past runs.
• Sub-tasks pay for their own steps. A run an agent starts as a sub-task draws credits for its own steps, the same as any other run.
• Dev runs don't draw credits. Their traces still show what each step would cost, but nothing comes off your balance, and Credits used counts Production only.

Where you see credits

• In the trace, each billable step shows a coin icon with its credits. Hover it to see, for example, 1 run credit charged for this step. A step with nested steps shows the total for its whole branch, and the tooltip splits the step's own credits from its nested steps' credits.
• In Executions, each run shows its credits under its cost, for example 5 credits.
• Settings → Usage shows what the whole workspace has used. See the next section.

Plans, allowance, and top-ups

Paid self-serve plans include a monthly credit allowance that resets each billing cycle. Your allowance is used first. Top-up credits are extra credits you buy on top of your plan. They roll over instead of resetting, and they're used once the monthly allowance runs out. On the Free plan, credits are a one-time grant rather than a monthly allowance.

Settings → Usage has a Credits used card:

• What you've used this cycle against your allowance, for example 4,232 of 10,000 credits this cycle, and what's left, with the reset date.
• Once the allowance is used up: Monthly allowance used — drawing on top-ups.
• A forecast of where the cycle will end at this week's pace. If a busy pace could use up the allowance before the reset, the forecast names the day.
• After a plan change, any credits left from your previous plan and the date they're usable until.
• View history, once there's more than a month of data, breaks usage down By source and By agent.
• Email admins as you near your monthly plan limit sends an email at 80% and 100% of the allowance.

On a new paid plan, the allowance appears once your first invoice is paid.

Settings → Billing shows your plan and a Top-up credits card. The card shows your balance, with Top up to buy more, Notify below to get an email when the balance runs low, and automatic top-ups on paid plans. Enterprise and self-hosted workspaces don't see the Billing page, because their terms are set by contract. Usage is still there.

To see how a typical run adds up, open What's a credit? on the Plans page.

When credits run out

When your balance reaches zero, new runs are refused before they start. An API caller gets HTTP 429 with the code CREDITS_EXHAUSTED:

You're out of credits. Top up or upgrade your plan to keep your agents running.

• Refused runs appear in Executions as Blocked, and a banner across Fruxon says you're out of credits.
• Fruxon emails the workspace's Owners and Admins when the balance runs out.
• Runs already under way are never cut off. They finish, and their steps are charged even past zero. The overdraw shows on the Top-up credits card as credits — overdrawn, and the next credits you get pay it off first.
• A running agent's sub-tasks aren't refused, so it can finish the work it started.
• Automation stages pause and try again later. See Spend limits and paused stages.

To avoid running out, turn on automatic top-ups or the low-balance alert on Settings → Billing.

Models Fruxon runs for you

Where Fruxon offers models it runs for you, you don't need an account with the model provider. Their token cost is paid from a prepaid Model usage balance instead of by your provider.

• Settings → Billing → Model usage balance shows what's left, with Add balance and automatic top-ups. Usage is charged at the provider's published rate plus a service fee. The card shows the fee percentage.
• Settings → Usage → Model usage lists each charge with its Provider cost, Service fee, and the total Charged.
• Dev runs on these models draw from the balance too, because the provider charges for test tokens like any others. Runs on these models also draw run credits for their steps, like any other run.
• Custom pricing doesn't change what's drawn from this balance. It only changes your cost reports and spend limits.

When the balance is empty, new runs on these models are refused with a message telling you to top up your balance or switch the agent to a provider configured with your own credentials. Runs already under way finish. Any usage past zero shows on the card as unpaid, and agents on these models stay paused until you top up.

These cards appear only where Fruxon offers models it runs for you.

Next steps

• Observability: traces, cost over time, and alerts
• Automations: how spend limits and credits pause automation stages
• Testing: test agents in the Dev bucket before production
• Settings: Usage, Billing, and Model pricing