> ## Documentation Index
> Fetch the complete documentation index at: https://docs.platform.toloka.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Pipeline pricing

> How Data Labeling pipeline runs are forecast, which cost lines apply, what Expected and Maximum mean, how deployment freezes a contract, and what canceled or failed items cost.

A pipeline run's cost is a forecast you approve before deployment, followed by pay-as-you-go charges for the items it processes. For the team balance, top-ups, automatic funding pauses, and the difference between pipeline and Training settlement, see [Billing](https://docs.platform.toloka.ai/docs/core/billing.md).

![The pricing model: the prices you set and the rates the platform sets both feed the Cost breakdown (forecast) window, which computes an Expected and a Maximum total; at deploy the contract freezes the per-item prices for the run; your balance is drawn down pay-as-you-go as items process, and an empty balance pauses the run until you top up and Resume.](https://mintcdn.com/toloka-platform/e2y6apyB8pyO00uJ/images/explanation/pricing-and-billing/pricing-model.svg?fit=max\&auto=format\&n=e2y6apyB8pyO00uJ\&q=85\&s=354fe6eb664ae53710393e5c4ac986a3)

## What you set and what the platform sets

You set the per-task price, the share of items sampled for human review, and the per-expert task cap. The platform sets generation token rates and review-call prices. Human-review pricing uses the per-task price you set, scaled by review time divided by labeling time, plus a 10% premium. For expert labeling, **Expert Service Cost** shows the per-task price plus the platform commission, displayed as 33%.

## The cost breakdown forecast

Open **Cost breakdown (forecast)** to compare **Price**, **Quantity**, **Expected**, and **Maximum** for each billable line and the run's **Total**. Rows are grouped by node; expand LLM review lines to inspect individual quality criteria. Invalid nodes appear in a separate collapsible section and are excluded from the totals.

Each line carries two totals, and so does the run:

* **Expected** is the average forecast, or the run's likely cost.
* **Maximum** is the run's cost if every item triggered every billable event. It is the forecast ceiling.

Billing is pay-as-you-go for the items a run actually processes. The actual total can come in below Expected. Billing commits you only to actual costs, while Maximum caps the forecast.

## The cost lines

Each billable line applies to one labeling path and is priced by one unit. The **Cost breakdown (forecast)** window lists the lines your pipeline actually incurs.

| Line | Applies to | Priced per |
| - | - | - |
| **Expert Service Cost** | Expert labeling | Each expert per item, at your per-task price plus 33% commission |
| **LLM QA** | Any node reviewed by an LLM | Each review call, with one call per quality criterion per labeled item at the review model's per-call price |
| **Human QA** | A Generation node with human review | Each sampled review, at a per-review price |
| **Onboarding (one-time per expert)** or **Training (one-time per expert)** | Expert labeling with an onboarding course; labeled Training when an [exam](https://docs.platform.toloka.ai/docs/data/reference/nodes/generation.md#exam) is enabled | Once per expert who joins the task, including human reviewers, spread across the batch |
| **Exam (one-time per expert)** | Expert labeling with an [exam](https://docs.platform.toloka.ai/docs/data/reference/nodes/generation.md#exam) | Once per expert who passes the exam and starts work, including human reviewers |
| **Video tool** | Any labeling whose LLM review watches a video field, expert or synthetic | One call per labeled answer, whatever the number of criteria, at a base rate plus a per-minute rate over every video watched. The criteria always bring in the video fields they reach through field links; synthetic labeling adds the video fields its outputs reach. A video field reached more than once counts once, and one that nothing links to is never charged. A video field with no declared maximum counts as five minutes |
| **Synthetic generation** | Synthetic labeling | Input, cached-input, and output tokens at the rates listed beneath this line, sharing one per-item cap of twice the generation budget per attempt. LLM QA is billed separately |

| Setting or event | Effect on the forecast |
| - | - |
| Expert count | Expert Service Cost scales with the number of experts per item |
| Quality criteria and repeating fields | LLM QA multiplies the criterion count by the relevant repetition maxima and the Quorum expert count |
| Quorum review retries | Maximum budgets twice the initial review-call count; other nodes add no extra review rounds |
| Human-review rework | An item can require more than one review |
| Onboarding and batch size | One-time charges are spread across the batch; with the per-expert cap off, their per-item share shrinks as the batch grows |

## The contract freeze

Deployment fixes the run's per-item prices for its lifetime, even if platform prices later change. Each clone receives a new estimate and contract when it deploys. Click **Contract** on the run to reopen its frozen breakdown.

## Pay from the team balance

Pipeline runs use the balance shared by every project in the team. The deploy window can recommend funding levels, and a run can pause automatically if it cannot reserve the next node's cost. [Billing](https://docs.platform.toloka.ai/docs/core/billing.md#pipeline-runs) covers the balance requirement, funding levels, automatic pause, top-up, and resume flow.

## What a canceled or failed item costs

When an item starts on a node, the platform reserves the node's maximum cost for the item from your balance. When the item completes or is canceled on the node, you pay for the work recorded on the item, and the rest of the reserved amount returns to your balance. For the reserved lines, you never pay more than the reserved amount. Human QA and the lines marked one-time per expert are not reserved. You pay for them when the node ends.

How the item ends on the node sets what you pay:

| The item on the node | Reserved lines | Human QA and one-time per-expert lines |
| - | - | - |
| Completes | The recorded work | The recorded work |
| Is canceled | The work recorded until the item ends on the node. A line with no recorded work costs nothing | The recorded work |
| Fails | Nothing. The reserved amount returns in full | The recorded work |

On a canceled item, each reserved line counts this work:

| Line | Generation node | Quorum node |
| - | - | - |
| **Expert Service Cost** | One task, if the task was accepted. A task accepted without review when you cancel counts | One task for each submitted answer that was accepted |
| **LLM QA** | The review calls made | The review calls made |
| **Video tool** | One call, if LLM QA checked the item | One call for each expert in the configured overlap, if LLM QA checked an answer |
| **Synthetic generation** | Input, cached-input, and output tokens used, including rejected generations, within the shared per-item cap | Does not apply |

For example, LLM QA rejects a Generation item, and you cancel the run before the expert's redo. The item pays for the LLM QA calls made, and for 1 Video tool call if its LLM QA watches a video. The rest of the reserved amount returns to your balance. An item canceled before any work is recorded costs nothing.

## How a Match node shapes the estimate

Routing an item through a Match node costs nothing on its own. Branching changes the totals: **Expected** assumes items spread evenly across the branches, while **Maximum** sends the full item count down the most expensive complete path. The exact rules live on [the Match node reference](https://docs.platform.toloka.ai/docs/data/reference/nodes/match.md). The forecast does not know your real match rates.

A Generation node with human review stops reviewing at the **Maximum** of its **Human QA** line: 3 reviews for each item of the run. This cap counts every item of the run, not only the node's even share. A branch that receives more items than its share still gets human review on them.

## Related

* [Billing](https://docs.platform.toloka.ai/docs/core/billing.md#pipeline-runs) — fund a pipeline run and recover from an automatic funding pause
* [Label a large dataset in batches](https://docs.platform.toloka.ai/docs/data/guides/label-at-volume.md) — use the cost window in the deploy workflow and size a monthly volume
* [Quality control](https://docs.platform.toloka.ai/docs/data/concepts/quality-control.md) — understand what review costs and why
* [Generation node](https://docs.platform.toloka.ai/docs/data/reference/nodes/generation.md) — configure the per-task price, review, and token settings that drive these lines
* [Quorum node](https://docs.platform.toloka.ai/docs/data/reference/nodes/quorum.md) — understand how the expert count scales the expert and review lines


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.