ETL & ELT Pricing Guide (2026): What Data Integration Actually Costs

The reason is structural: each vendor meters a different unit — rows, events, credits, compute-hours — and the unit that looks cheapest on the pricing page is rarely the one that stays cheap under your actual workload.

This guide explains the five pricing models in use across the data integration market in 2026, breaks down what the major tools actually charge today, and covers the cost traps we see when reviewing client pipelines. Prices below were verified in July 2026. Vendors change pricing frequently; treat every number as a planning figure and confirm on the vendor's pricing page before signing.

The five pricing models

1. Monthly Active Rows (MAR). You pay for distinct rows inserted, updated, or deleted per month. A row that changes ten times in a month counts once. Fivetran is the reference implementation. Predictable for stable sources, volatile for high-churn tables.

2. Event-based. Every record written to the destination counts as one event, including repeat updates to the same row. Hevo uses this model. Simpler to reason about than MAR, but repeated updates to the same records cost more than they would under MAR.

3. Row-based tiers. Flat monthly fee for a row allowance, stepping up in tiers. Stitch built its business on this model. The most budget-predictable of the usage models, as long as you stay inside your tier.

4. Credit or capacity-based. You buy credits consumed by sync volume, or provision fixed capacity (workers) regardless of volume. Airbyte Cloud uses credits on its entry plan and capacity-based Data Workers on higher plans.

5. Compute-hours. You pay for the infrastructure that runs the job, not the data that moves through it. AWS Glue charges per DPU-hour. Cost tracks job runtime and cluster size, which puts pricing in your engineers' hands — for better and worse.

There is a sixth option that isn't a pricing model: open-source self-hosted (Airbyte OSS, Meltano, Singer taps). The software is free. The infrastructure and the engineering time are not, and they usually dominate the real cost. More on that in the buy-vs-build section.

Tool-by-tool breakdown

Fivetran

Fivetran prices on MAR with per-connection consumption curves: each connection is billed separately, and the per-MAR unit cost declines as that connection's volume grows.

Current structure:

• Free: 500,000 MAR/month, plus small allowances for activations and model runs. An evaluation tier; a single moderately busy production database exceeds it.
• Standard: usage-based, roughly $500 per million MAR at entry volumes, 15-minute minimum sync frequency, 700+ connectors.
• Enterprise: roughly a third more per MAR, 1-minute syncs, enterprise database connectors (Oracle, SAP, Db2), 99.9% SLA.
• Business Critical: roughly double Standard's rate, adds customer-managed keys, PCI DSS Level 1, private networking.

Three 2025–2026 changes materially raised bills for many teams:

1. Connector-level MAR billing (March 2025). MAR is calculated per connection, not pooled across the account. Teams running many mid-size connectors lost their volume discounts; increases of 40–70% were common for multi-connector setups.
2. Deletes now count as paid MAR (January 2026). High-churn tables — orders, sessions, inventory — got more expensive overnight.
3. $5 base charge per connection generating 1–1M MAR (January 2026). Trivial for one connector, real money for a fleet of small ones.

What stays free: initial historical syncs and re-syncs of unchanged rows.

Realistic budget: a mid-market company replicating 10–15 sources into Snowflake typically lands around $2,000–$4,000/month, with the first month higher if backfills trigger. Annual contracts discount roughly 15%, two-year around 20%, and volume tiers discount further past ~$36K annual spend.

Best fit: teams with few, large, stable sources that value connector reliability over cost, and have the discipline to monitor MAR per connection.

Hevo

Hevo prices on events: every record inserted, updated, or deleted in the destination counts once, every time.

• Free: 1M events/month, limited connector set.
• Starter: from $239/month (annual billing) including 5M events; a slider scales the quota and price. 150+ connectors, up to 10 users.
• Professional: from $679/month including 20M events, unlimited users, streaming-grade sync options.
• Business Critical: custom pricing; where HIPAA, RBAC, and SSO live.

Overages are billed as on-demand events on top of the plan, and Hevo provides a buffer so pipelines don't stop mid-month. New sources get 14 days of free events, which makes proofs-of-concept cheap.

The event model's trap is repeat updates: a row that updates hourly costs 720 events a month under Hevo where Fivetran's MAR would count it once. For CDC-heavy operational sources, model this before assuming Hevo is cheaper. For insert-heavy workloads, it often genuinely is.

Best fit: analytics teams with insert-dominant workloads that want a fixed monthly number and strong support without enterprise procurement.

Stitch

Stitch is row-based: flat tiers from roughly $100/month for 5M rows up to about $1,250/month on Standard, with a Premium tier around $3,000/month for 1B rows and HIPAA compliance. Around 130–140 connectors, Singer-based.

The number that matters most in 2026 is strategic, not financial: Stitch is now part of Qlik and is being folded into Qlik Talend Cloud. The Stitch site directs new signups toward Qlik Talend Cloud, and standalone feature development has effectively ended. Existing customers still run fine today. New teams should treat Stitch as a sunset product and price in a future migration, not just a monthly fee.

Best fit: existing customers with stable pipelines, and teams already inside the Qlik ecosystem. We would not start a new greenfield deployment on standalone Stitch in 2026.

Airbyte

Airbyte spans the widest cost range in the market because it's really three products:

• Core (open source, self-hosted): free software, unlimited volume, 600+ connectors. Real cost: production-grade Kubernetes infrastructure typically runs $500–$3,000+/month, plus a commonly cited 20–40 engineering hours per month for upgrades, connector issues, and pipeline maintenance.
• Cloud Standard: from a $10/month minimum, credit-based (around $2.50 per credit), with volume-based consumption. Cheap entry, but failed syncs still consume credits, and large initial backfills can generate one-time charges in the thousands.
• Plus / Pro / Enterprise Flex: capacity-based Data Worker pricing, entry around $25,000/year. SSO, RBAC, and enterprise connectors (SAP, NetSuite, Workday, Oracle) are gated here.

The structural issue is the gap: there is very little between $10/month self-serve and $25K/year contracts. Growing teams hit that cliff mid-scale. The structural advantage: at high volumes, self-hosted Airbyte with competent ops is the cheapest per-terabyte option on this list by a wide margin.

Best fit: engineering-led teams that either stay small on Cloud Standard or have genuine platform capacity to run self-hosted. Wrong fit: teams with no one to own the infrastructure.

AWS Glue

Glue prices compute, not data:

• ETL jobs and interactive sessions: $0.44 per DPU-hour, billed per second, 1-minute minimum. A DPU is 4 vCPU and 16 GB memory. Flex execution drops non-urgent batch work to $0.29 per DPU-hour, about a third cheaper.
• Crawlers: $0.44 per DPU-hour, 10-minute minimum.
• Data Catalog: first million objects and million requests free monthly, then $1 per 100,000 objects.
• DataBrew: $1 per 30-minute interactive session.
• Zero-ETL integrations: ingestion billed around $1.50/GB, plus destination costs.

Worked example, straight from the model: a Spark job using 6 DPUs for 15 minutes costs 6 × 0.25 × $0.44 = $0.66. Sounds like pocket change. The bill grows through frequency (hourly jobs run 720 times a month), over-provisioned DPUs, and forgotten development endpoints, which have no auto-timeout and keep billing until someone kills them.

Best fit: AWS-native teams doing heavy transformation work who can engineer their own cost efficiency. Glue is not really a connector product; comparing it to Fivetran is comparing an engine to a car.

Comparison table

The cost traps

Connector sprawl. Every "quick" new source adds a fixed floor: a base charge on Fivetran, a tier bump on Stitch, another pipeline to maintain anywhere. Audit quarterly; most stacks we review carry connectors nobody has queried in months.

Backfill and re-sync spikes. Initial syncs are free on Fivetran but not everywhere; on credit-based models a schema change that forces a full re-sync of a large table can produce a four-figure one-time charge. Know your tool's re-sync billing before you approve a schema migration.

Churn-heavy tables. Since deletes became billable MAR, tables with heavy insert-delete cycles (queues, sessions, staging tables) are disproportionately expensive to replicate. Often the fix is not a cheaper tool but a question: does this table belong in the warehouse at all?

Sync frequency defaults. The difference between 5-minute and 6-hour syncs on a low-priority source is real money every month, and most sources were set to a frequency once and never revisited.

Destination costs. Warehouse compute and storage — Snowflake credits, BigQuery scans, Redshift nodes — are a separate bill that ingestion decisions directly drive. Third-party analyses regularly find destination costs equal to or exceeding the pipeline tool's fee. Budget them together.

Egress and networking. Cross-region traffic, NAT gateways for self-hosted deployments, and private networking surcharges sit outside every pricing page and inside every invoice.

A total-cost worksheet

Before comparing vendors, write down for your own workload:

1. Number of sources, and for each: rows changed per month, records written per month, and update frequency (these three numbers price you under MAR, events, and rows respectively)
2. Highest-churn tables, including deletes
3. Required sync frequency per source — actual business need, not default
4. Any table over ~100M rows that might ever need a full re-sync
5. Compliance requirements (HIPAA/PCI push you into top tiers everywhere)
6. Destination compute cost per additional TB ingested
7. For self-hosted: who owns the infrastructure, and what their hours cost

Run those numbers through two or three vendors' models. The ranking that comes out is frequently the opposite of the pricing-page ranking.

Buy vs build

Self-hosting or hand-rolling pipelines wins on paper and loses in payroll. The realistic break-even: if your monthly managed-tool bill is below what 25% of a data engineer costs you, buy. Self-hosted Airbyte or custom pipelines make sense when volumes are large enough that managed per-row pricing exceeds a platform engineer's time, and — this is the condition teams skip — when that engineer actually exists and stays. Pipelines built by someone who then leaves are how we meet many of our clients. If that's the situation you're in, our database handover checklist is the place to start.

FAQ

Which ETL tools offer the best pricing models for growing businesses?

For predictability while growing: Hevo's event tiers or Stitch-style row tiers, because the monthly number is known in advance. For lowest cost at small scale: Airbyte Cloud's $10 entry. For growth into large, stable data volumes: Fivetran's per-connection curves reward consolidation into fewer, bigger connectors. The honest answer is workload-dependent: insert-heavy favors events, churn-heavy favors MAR, and very high volume favors self-hosting with real ops capacity.

How much does ETL cost in 2026?

A small analytics stack (3–5 sources, modest volume) runs $0–$300/month on free tiers or entry plans. A mid-market deployment of 10–15 sources typically lands between $1,000 and $4,000/month on managed platforms, before destination warehouse costs. Enterprise deployments with compliance requirements run $10,000/month and up. Self-hosted open source trades those fees for $500–$3,000/month in infrastructure plus meaningful engineering time.

What is the difference between ETL pricing models?

MAR (Fivetran) counts distinct rows changed per month, once each. Event pricing (Hevo) counts every record write, including repeated updates to the same row. Row tiers (Stitch) sell a flat allowance of replicated rows. Credit and capacity models (Airbyte Cloud) meter sync volume or provisioned workers. Compute pricing (AWS Glue) bills the runtime of the jobs themselves. The same workload can differ in cost by 3–5x across these models, which is why the model matters more than the sticker price.

How do ETL tools vary in cost as data grows?

Usage models (MAR, events, credits) scale linearly or worse with data change volume, with unit discounts partially offsetting. Tiered models step up in jumps and become expensive at tier boundaries. Compute models scale with job runtime, which good engineering can hold nearly flat. Self-hosted costs scale in steps (bigger cluster, more ops time). Teams that grow fastest usually migrate models at least once: entry tool to scale tool around the point where the monthly bill crosses a senior engineer's day rate.

Pricing shifts like Fivetran's 2026 changes are often the trigger for a platform migration, and migrations are where pipelines break. If you're planning a move between ETL tools or re-platforming a warehouse, our migration support playbook covers the sequence, and our data operations team handles migrations end to end. Describe your situation through the contact form and a senior lead will respond within one business day.