Ingestion is the least visible line on most data budgets until the invoice changes. Fivetran has changed how it counts billable usage twice since March 2025, and on June 1, 2026 it completed its merger with dbt Labs (Fivetran, 2026). Teams buying pipeline tools now choose between three pricing units: rows changed, gigabytes moved and reserved capacity. Each one rewards a particular data shape and penalises another.
This explainer covers how each model counts, where costs build up, and what to check before signing. Prices below are as listed in vendor documentation as of October 2026. They change often, so confirm them against the current page.
What is being priced
A managed ingestion tool copies data from SaaS applications, production databases and files into a warehouse or lakehouse. It handles authentication, API pagination, schema changes and incremental syncs, so the data team does not maintain that code.
The vendor's workload scales with several variables at once: how many sources are connected, how often they sync, how many records change between syncs and how large those records are. No single unit captures all four. Each pricing model picks one proxy and accepts that it will be unfair to some customers.
Model one: monthly active rows
Fivetran bills on monthly active rows (MAR). A row counts once per calendar month when it is inserted, updated or deleted, however many times it changes after that (Fivetran, 2026). Fivetran's documentation also explains that a new connection's initial historical sync does not add paid MAR (Fivetran, n.d.).
Two changes reshaped the model. On March 1, 2025, Fivetran moved volume tiering from the account to the individual connection. Its documentation says the change was calibrated to be roughly neutral for a typical user, but that customers with many equal-sized connections would see higher prices and customers with one dominant connection would see lower ones (Fivetran, 2025). Contracts signed before that date keep account-level tiering until renewal.
The second change took effect on January 1, 2026 for monthly pay-as-you-go accounts, and applies to annual contracts at renewal. Deletes now count toward paid MAR alongside inserts and updates. Every standard connection that generates between 1 and 1 million MAR in a month also carries a minimum charge, which Fivetran's documentation illustrates as $5 on the Standard plan (Fivetran, 2026). Connections with zero paid MAR are exempt, as is the Free plan.
MAR suits sources where a small share of a large table changes each month. In a hypothetical 50 million row orders table where 200,000 rows change in a month, only those 200,000 rows bill. The model is harder on tables where most rows change constantly, such as status or event tables, and connection-level tiering is harder on estates with many small connectors.
Model two: volume moved
Estuary prices on gigabytes plus connectors. Its pricing page lists $0.50 per GB of data moved, $100 per month for each of the first six connector instances and $50 per month for each additional instance (Estuary, 2026). Because a pipeline both captures from a source and delivers to a destination, Estuary's own materials describe a single source-to-destination pipeline as $1 per GB in total.
Airbyte Cloud's self-serve Standard and Plus plans use credits with a split rate: 6 credits per million rows for API sources, and 4 credits per GB for database and file sources (Airbyte, 2026). The split reflects a real difference. API sources tend to return small records slowly, while database replication moves large volumes at low cost per row.
Volume pricing is easier to forecast from metrics a team already has, such as replication log size or file sizes in object storage. It penalises wide rows, verbose JSON payloads and full-table reloads, because every byte bills whether or not the business considers it a change.
Model three: reserved capacity
Airbyte's Pro plan (formerly Teams) and Enterprise Flex replace volume billing with capacity measured in Data Workers. Airbyte describes a Data Worker as a unit of pipeline compute that typically runs about three syncs concurrently, and says the number a customer needs depends on how much data must move at the same time rather than how much moves in total (Airbyte, 2025).
Capacity pricing behaves like reserved warehouse compute. Cost is predictable and does not jump when a source suddenly emits more changes. The trade is utilisation risk. A team that schedules every sync at 2am needs more workers than one that spreads syncs across the day, and idle capacity is paid for regardless.
Self-hosting open-source connectors is the limiting case of this model. The licence cost is zero, and the bill moves to cloud compute and engineering time, which rarely sit on the same budget line as the tool they replaced.
What to check before buying
- Price your real change profile. Pull one month of change counts and byte volumes per source from database logs or API metadata, then price that same month under each model. List rates compared side by side say little on their own.
- Count connectors, not only volume. Connection-level tiering and per-connection minimums mean twenty small sources can cost more than one large source moving the same total data.
- Get re-sync and schema-change billing in writing. A full re-sync after a schema change, or a backfill after an outage, can be the most expensive event of the year. Ask the vendor how each is counted.
- Check which rules your contract is on. Fivetran's 2025 and 2026 changes reach annual customers at renewal, so two customers on the same plan can be billed under different rules.
- Include the destination. Every ingested change also consumes merge compute in the warehouse. A connector that syncs more often may cost less itself and still raise the warehouse bill.
What commonly goes wrong
The most common failure is a forecast built on an average month. Ingestion bills are driven by unusual months: a migration that rewrites every row, an application that updates a timestamp column nightly, a new connector added by an analyst without review.
The second is evaluating the ingestion vendor in isolation. Fivetran and dbt Labs now operate as one company, with George Fraser as CEO and Tristan Handy as president, and their first joint releases include dbt Core v2.0 under an Apache 2.0 licence (Fivetran, 2026; TechTarget, 2026). Teams renewing either product should ask whether the two will be priced together, and on which unit.
The third is watching nothing until the invoice arrives. Fivetran documents per-connection usage monitoring (Fivetran, n.d.). Whichever vendor a team picks, a weekly check of usage by connection catches a runaway source before month end.
Sources
- Fivetran. "Usage-Based Pricing Updates 2026." 2026. https://fivetran.com/docs/core-concepts/usage-based-pricing/pricing-updates/2026-pricing-updates
- Fivetran. "Usage-Based Pricing Updates 2025." 2025. https://fivetran.com/docs/usage-based-pricing/pricing-updates/2025-pricing-faq
- Fivetran. "Pre-March 2025 Usage and Pricing." 2025. https://fivetran.com/docs/usage-based-pricing/pricing-updates/pre-march-2025-pricing
- Fivetran. "Why Has My New Connection Not Contributed to Paid MAR?" n.d. https://fivetran.com/docs/core-concepts/usage-based-pricing/troubleshooting/new-connector-no-paid-mar
- Fivetran. "Monitor and Optimize Usage." n.d. https://fivetran.com/docs/usage-based-pricing/manage-mar
- Estuary. "Pricing & Plans." 2026. https://estuary.dev/pricing/
- Airbyte. "Manage billing and credits." 2026. https://docs.airbyte.com/platform/cloud/managing-airbyte-cloud/manage-credits
- Airbyte. "Airbyte 2.0: The Data Platform Built for the AI Era." 2025. https://airbyte.com/blog/airbyte-2-0
- Airbyte. "Pricing." https://airbyte.com/pricing
- Fivetran. "Fivetran, dbt Labs Complete Merger to Create the Data Infrastructure for Trusted AI Agents." 2026. https://www.fivetran.com/press/fivetran-dbt-labs-complete-merger-to-create-the-data-infrastructure-for-trusted-ai-agents
- TechTarget. "Fivetran, dbt Labs complete merger to form data layer for AI." 2026. https://www.techtarget.com/data-technologies/news/366643590/Fivetran-DBT-Labs-complete-merger-to-form-data-layer-for-AI