Skip to content
Developer Tools · Workflow Orchestration

Top 6 Workflow Orchestration Platforms for 2026: Temporal vs Airflow vs Prefect vs Dagster vs Kestra vs Windmill

Data-pipeline orchestration and durable execution are two different purchases. Priced separately, with the Prefect acquisition of Dagster accounted for.

By ·Sep 18, 2026·18 min·6 tools compared
Workflow OrchestrationDurable ExecutionTemporalApache AirflowData EngineeringDeveloper Tools

Buyers conflate two different products here, and the conflation costs real money, so this page starts by splitting them.

Data-pipeline orchestration schedules and sequences work, usually on a timetable, usually over data. Read twelve sources, transform, write a warehouse table, alert on failure. That is Apache Airflow, Prefect, Dagster, and Kestra.

Durable execution persists the state of a running program so it resumes exactly where it stopped after a crash, a redeploy, or a three-week wait for a human. Charge the card, provision the account, wait for identity verification, compensate if the third call fails. That is Temporal.

A useful test: if your workflow has a schedule, you want orchestration. If it has a trigger and a long life, you want durable execution. Both categories describe themselves as workflow engines, which is why teams routinely buy the wrong one and then blame the tool.

The short answer. Temporal for application and service workflows that must not lose their place. Apache Airflow for scheduled batch pipelines across heterogeneous systems, evaluated against Airflow 3 rather than the 2.x you may remember. Prefect for Python-first, dynamic, data-dependent workflows. Dagster for greenfield data platforms where lineage and freshness are requirements. Kestra when the Python requirement is the problem. Windmill when the real target is a pile of untracked cron jobs rather than a data platform.

Note

Two market changes this year reshape the shortlist. Prefect agreed to acquire Dagster Labs on 13 July 2026, so the two leading Airflow successors now have one owner; both products are stated to continue independently with unchanged pricing. And Airflow 3.3.0 shipped on 6 July 2026 with a Language Task SDK for writing tasks in Java and Go, which removes one of the standing arguments for leaving Airflow.

On open source, this page refuses the word free. Airflow's licence costs nothing and a production deployment is a scheduler, a triggerer, a webserver, a worker pool, and a Postgres metadata database, all needing monitoring, capacity planning, and upgrades that periodically break DAGs. Realistically that is half a platform engineer as a standing commitment. Self-hosting Temporal is harder still. Every open-source entry below states the operational burden rather than a zero.

Every price here was read from the vendor's own pricing page in September 2026, and licences were checked against each project's repository. Where a vendor publishes no rate card, such as Kestra Enterprise and Kestra Cloud, this page says so.

Orchestration is the scheduling layer, not the connector layer. If your question is which tool moves data between systems and which one models it, that is a separate purchase covered in the data pipeline and ETL tools comparison, which this page deliberately does not repeat. For the CI layer that deploys these workflows, see the CI/CD platform comparison.

Quick Comparison

PlatformWhich ProblemLicenceFree TierPublished Price
TemporalDurable execution for application and service workflowsMIT (server), self-hostable$150 in Temporal Cloud credits, 90 daysFrom $50 per million Actions, down to $25 at volume
Apache AirflowScheduled data-pipeline orchestration, broadest ecosystemApache 2.0Unlimited, self-hosted, no vendor tier$0 licence; Astro deployments from $0.35/hr, workers from $0.13/hr
PrefectPythonic data and ML orchestration with dynamic workflowsApache 2.0Hobby: 2 users, 1 workspace, 5 deployments$100/mo Starter, $100/user/mo Team
DagsterAsset-based data platforms where lineage is the modelApache 2.030-day trial only$10/mo + $0.040/credit Solo, $100/mo + $0.035/credit Starter
KestraDeclarative, language-agnostic orchestration without Python lock-inApache 2.0OSS edition, unlimited flows and executionsEnterprise not published; Kestra Cloud is usage-based
WindmillTurning scripts into internal workflows, apps, and endpointsApache 2.0 / AGPLv3 / proprietary enterprise featuresSelf-hosted: 3 workspaces, 50 users, 10 GiB storageCloud Pro from $120/mo, $20/developer seat, $50/worker/mo

Temporal

Which Problem
Durable execution for application and service workflows
Licence
MIT (server), self-hostable
Free Tier
$150 in Temporal Cloud credits, 90 days
Published Price
From $50 per million Actions, down to $25 at volume

Apache Airflow

Which Problem
Scheduled data-pipeline orchestration, broadest ecosystem
Licence
Apache 2.0
Free Tier
Unlimited, self-hosted, no vendor tier
Published Price
$0 licence; Astro deployments from $0.35/hr, workers from $0.13/hr

Prefect

Which Problem
Pythonic data and ML orchestration with dynamic workflows
Licence
Apache 2.0
Free Tier
Hobby: 2 users, 1 workspace, 5 deployments
Published Price
$100/mo Starter, $100/user/mo Team

Dagster

Which Problem
Asset-based data platforms where lineage is the model
Licence
Apache 2.0
Free Tier
30-day trial only
Published Price
$10/mo + $0.040/credit Solo, $100/mo + $0.035/credit Starter

Kestra

Which Problem
Declarative, language-agnostic orchestration without Python lock-in
Licence
Apache 2.0
Free Tier
OSS edition, unlimited flows and executions
Published Price
Enterprise not published; Kestra Cloud is usage-based

Windmill

Which Problem
Turning scripts into internal workflows, apps, and endpoints
Licence
Apache 2.0 / AGPLv3 / proprietary enterprise features
Free Tier
Self-hosted: 3 workspaces, 50 users, 10 GiB storage
Published Price
Cloud Pro from $120/mo, $20/developer seat, $50/worker/mo
1

Temporal

Best Overall

Best for: Application and service workflows that must survive process crashes, deploys, and multi-day waits, such as payment flows, order fulfilment, provisioning, and long-running agent loops

Temporal is not competing with Airflow and buyers who treat it as an alternative end up disappointed in both directions. It is a durable execution engine: your workflow is ordinary code, its state is persisted after every step, and it resumes exactly where it stopped after a crash, a deploy, or a three-week wait for a human. If your problem is a distributed saga that must not lose its place, nothing else on this page is the right purchase.

Pros

  • Workflows are written in ordinary application code (Go, Java, TypeScript, Python, .NET, PHP, Ruby) rather than a DAG DSL, so business logic stays testable
  • Durable state means a workflow survives worker crashes, redeploys, and waits measured in weeks without bespoke checkpointing code
  • Server is MIT licensed and genuinely self-hostable, so Temporal Cloud is a convenience purchase rather than the only path
  • Temporal Cloud states no features are locked behind plan upgrades, with tiers differing on support and SLA rather than capability
  • Usage-based Action pricing scales down to real zero for low-volume workflows, unlike per-seat orchestration products

Cons

  • Action-based billing ($50 per million Actions, down to $25 at volume) is hard to forecast before you have written the workflows, because an Action is a unit of workflow progress rather than a run
  • Active storage at $0.042 per GB-hour is a separate and easily underestimated line item for workflows with large payloads or long retention
  • Self-hosting a production Temporal cluster is genuinely demanding: it needs Cassandra or a supported database, careful capacity planning, and real operational knowledge
  • The programming model requires determinism in workflow code, which is a real constraint that trips up teams for the first few months
  • Poor fit for scheduled batch data pipelines, where Airflow-style tooling is simpler and better supported
Honest Weakness: Temporal's pricing is the hardest to forecast on this page and its learning curve is the steepest. Actions are units of workflow progress, not workflow runs, so a chatty workflow with many activities and timers costs far more than the run count suggests, and nobody can model that accurately before writing the code. Combine that with the determinism constraint in workflow code and the honest advice is to prototype on the $150 free credit specifically to measure Action consumption, before committing anything to a budget.

Durable Execution Is Not Orchestration

The clearest way to tell which half of this page you need: does your workflow have a schedule, or does it have a trigger and a long life? A nightly job that reads twelve sources and writes a warehouse table is orchestration, and Airflow, Prefect, Dagster, or Kestra are the answer. A customer signup that provisions an account, calls three payment APIs with retries and compensation, and waits two days for identity verification is durable execution, and Temporal is the answer. Buyers conflate these because both are described as workflow engines, and the resulting mismatch is expensive in both directions.

What An Action Actually Is

Temporal bills per Action, and Actions are not workflow executions. Starting a workflow is an Action, each activity execution is an Action, each timer, each signal, each query, each child workflow start. A workflow with twenty activities and several retries can easily consume thirty or more Actions. At $50 per million that is still cheap for most workloads, but the forecasting error is usually one order of magnitude, not a few percent. Measure it on the free credit before you budget it.

Temporal Cloud free: $150 in credits, valid 90 days, no commitment. Pay as you go: Actions from $50 per million, falling to $25 per million at volume; Active Storage $0.042 per GB-hour; Retained Storage $0.00105 per GB-hour; Developer support included at 10 percent of usage. Business support: from $500 per month (the greater of $500 or 10 percent of usage), including 2.5M Actions. Enterprise and Mission Critical: custom. Self-hosted server: free, MIT licensed. As published on Temporal's pricing page, September 2026.

Visit Temporal
2

Apache Airflow

Best Open Source

Best for: Scheduled batch data pipelines in heterogeneous environments, where the number of systems you must talk to matters more than the elegance of the programming model

Airflow is still the default for scheduled data-pipeline orchestration and still the one with the most providers, the most hiring pool, and the most Stack Overflow answers. Airflow 3 (GA April 2025) fixed a great deal of what people complained about, and 3.3.0 landed in July 2026 with a Task SDK for writing tasks in Java and Go. It remains operationally heavy to self-host, which is the entire business model of Astronomer and the managed cloud offerings.

Pros

  • Apache 2.0 with the deepest provider ecosystem in the category, covering essentially every database, cloud service, and SaaS tool a pipeline needs to reach
  • Largest talent pool of any orchestrator, so hiring and handover are materially easier than with the newer alternatives
  • Airflow 3 delivered a genuine architecture refresh, including a new task execution interface, and 3.3.0 (July 2026) added a Language Task SDK for Java and Go plus pluggable retry policies
  • Managed options exist from multiple vendors, so a self-hosted deployment can be migrated to a managed one without changing the model
  • Asset and dataset-driven scheduling closes much of the gap with Dagster's asset model for teams that want it

Cons

  • Self-hosting is the real cost: a scheduler, webserver, triggerer, workers, a metadata database, and an executor choice, all of which need ongoing capacity and upgrade work
  • Dynamic workflows are still awkward compared with Prefect, because the DAG structure wants to be known ahead of time
  • Local development and testing remain weaker than the newer tools, which is a daily productivity tax rather than a one-off
  • Managed Airflow pricing is usage-based and multi-dimensional: on Astro, deployments start at $0.35 per hour on Developer and $0.42 on Team, workers from $0.13 per hour, and dedicated clusters from $2.40 per hour
  • Not suitable for durable application workflows, despite frequently being asked to do that job
Honest Weakness: Airflow's licence is free and its operational burden is the largest on this page. A production deployment is at minimum a scheduler, a triggerer, a webserver, a worker pool, and a Postgres metadata database, all of which need monitoring, capacity planning, and version upgrades that periodically break DAGs. The honest comparison against Astronomer or another managed Airflow is not $0 versus hourly rates. It is a fraction of a platform engineer versus hourly rates, and that fraction is rarely smaller than half a person.

Airflow 3 Changed The Calculus

A lot of Airflow criticism dates from the 2.x era and is repeated without checking. Airflow 3.0 reached GA on 22 April 2025 with a reworked task execution architecture. 3.1.0 (September 2025) added human-in-the-loop workflows and deadline alerts. 3.2.0 (April 2026) added asset partitioning and multi-team deployments. 3.3.0 (6 July 2026) added a state store for tasks and assets, a Language Task SDK for Java and Go, and pluggable retry policies. If your last evaluation of Airflow was against Airflow 2, it is out of date.

Apache Airflow: free, Apache 2.0, self-hosted, no limits. Real cost is scheduler, triggerer, webserver, worker, and metadata database infrastructure plus the engineering time to run them. Astronomer Astro (managed): usage-based, deployments from $0.35/hour on Developer and $0.42/hour on Team, workers from $0.13/hour billed only while tasks run, dedicated clusters from $2.40/hour on Team and above, additional ephemeral storage $0.0002 per GiB-hour. Business, Enterprise, and Private Cloud tiers are quote-only. As published on Astronomer's pricing page, September 2026.

Visit Apache Airflow
3

Prefect

Best Value

Best for: Python-first data and ML teams whose workflows are dynamic, where the shape of the DAG depends on data discovered at runtime rather than being known in advance

Prefect's core argument has always been that workflows are just Python functions with decorators, and that the DAG should be allowed to be dynamic. That is a real advantage over Airflow for branching, fan-out, and data-dependent workflows. In July 2026 Prefect acquired Dagster Labs, making it the owner of the two most adopted Airflow alternatives, with both products stated to continue independently and both price lists unchanged.

Pros

  • Workflows are plain Python with @flow and @task decorators, so local development, testing, and debugging work like normal code
  • Dynamic DAGs are native rather than a workaround, which suits ML and data workflows whose structure depends on runtime data
  • Hobby tier is genuinely free forever: 2 users, 1 workspace, 5 deployments, 500 serverless credit minutes a month, 7-day run retention
  • Bring-your-own-compute is supported from the Starter tier, so the cloud product can be a control plane over your own infrastructure
  • Open-source Prefect is Apache 2.0 and self-hostable, so the cloud is a convenience rather than a requirement

Cons

  • Python-only, which rules it out for teams whose transformation logic is Java, Go, or SQL-first
  • Team pricing at $100 per user per month is the highest per-seat number on this page and applies from 4 to 8 users
  • Starter at $100 a month covers only 3 users and 20 deployments, so the jump from Hobby to a working team setup is steep
  • Serverless credits are metered separately (500 minutes on Hobby, 75 hours on Starter, 225 hours on Team), so compute is a second dimension on the bill
  • SSO, RBAC, multiple workspaces, and directory sync are Enterprise-only, which is late for organisations with a security review
Honest Weakness: Prefect's free tier is genuinely useful and its first paid step is disproportionate. A four-person team goes from $0 to $400 a month, because Team is priced per user at $100. That is more per seat than any orchestrator here and it is charged on top of metered serverless credits. Self-hosting open-source Prefect avoids it entirely, which is a legitimate path, and it also means you are back to running a database and a server yourself.

Prefect Now Owns Dagster

On 13 July 2026 Prefect announced an agreement to acquire Dagster Labs, bringing the two most widely adopted Airflow successors under one owner, with the combined company expected to operate under the Prefect name from August 2026. Both companies state that Dagster and Dagster+ keep their names, open-source licence, and roadmap, that both products continue to be independently supported, and that pricing for both Prefect Cloud and Dagster+ is unchanged. That is the right commitment to make and it is also the commitment every acquirer makes. Treat long-term independence as a question to revisit at renewal rather than a settled fact.

Hobby: free forever, 2 users, 1 workspace, 5 deployments, 500 serverless credit minutes/month, 7-day run retention. Starter: $100/month, 3 users, 20 deployments, 75 serverless hours/month, bring-your-own compute. Team: $100/user/month, 4 to 8 users, 100 deployments, 225 serverless hours/month, service accounts, 24-hour audit log. Enterprise: custom, unlimited users and deployments, SSO, RBAC, multiple workspaces, 99.99 percent SLA. Prefect open source: free, Apache 2.0. As published on Prefect's pricing page, September 2026.

Visit Prefect
4

Dagster

Runner Up

Best for: Greenfield data platforms where the unit of work is the data asset rather than the task, and where lineage and freshness are first-class requirements rather than reporting features

Dagster's insight is that data teams do not care about tasks, they care about tables. Declaring assets rather than steps gives you lineage, freshness policies, and a catalogue for free, and it is the best model in this category for a platform built from scratch. Its pricing is credit-based and updated effective 1 May 2026. Its corporate situation changed in July 2026 when Prefect agreed to acquire Dagster Labs, with both products stated to continue.

Pros

  • Software-defined assets make lineage, freshness, and dependency tracking intrinsic rather than bolted on, which is a materially better model for data platforms
  • Excellent local development and testing story, which is the day-to-day difference most engineers notice against Airflow
  • Apache 2.0 open source with a self-hosted path, and Dagster states the open-source licence and roadmap are unchanged after the acquisition
  • Solo tier at $10 a month plus credits is the cheapest entry to a managed orchestrator on this page
  • 30-day free trial across all plans, and serverless compute priced transparently at $0.010 per minute

Cons

  • Credit-based billing is unpredictable by design: a credit is the sum of asset materialisations and ops executed, so a wide fan-out job consumes credits fast
  • No free tier at all, only a 30-day trial, which is weaker than Prefect's permanent Hobby plan
  • Solo is limited to 1 user, 1 code location, and 1 deployment; Starter allows 3 users and 5 code locations, so the caps bite early
  • The asset model is a genuine paradigm shift and porting existing task-based pipelines to it is a rewrite, not a migration
  • Pro pricing is quote-only, so the tier most organisations need has no published number
Honest Weakness: Dagster's asset model is the best idea in this category and it is also why Dagster is a poor migration target. If you already have hundreds of task-based DAGs, translating them into software-defined assets is a rewrite that produces no new business capability, and it usually stalls half-finished. Dagster earns its place on greenfield platforms. On an established Airflow estate, the honest recommendation is usually to stay on Airflow 3 and adopt its asset-driven scheduling instead.

Assets Versus Tasks

An Airflow DAG says run this, then run that. A Dagster definition says this table exists, it is derived from these two tables, and it should be no more than six hours stale. The second form gives you a lineage graph and a freshness contract without writing either. It also means the mental model, the testing strategy, and the debugging workflow are all different, which is why the migration cost is real and why the recommendation splits so cleanly between greenfield and established estates.

Solo: $10/month plus $0.040 per credit (pricing effective 1 May 2026), 1 user, 1 code location, 1 deployment. Starter: $100/month plus $0.035 per credit, up to 3 users, 5 code locations, 1 deployment, catalogue search and RBAC. Pro: custom, unlimited code locations and deployments, cost tracking, uptime SLAs. Serverless compute: $0.010 per minute. A credit is defined as the sum of asset materialisations and ops executed. 30-day free trial on all plans. Dagster open source: free, Apache 2.0. As published on Dagster's pricing page, September 2026.

Visit Dagster
5

Kestra

Best Free Option

Best for: Polyglot teams that want declarative YAML workflows with no Python requirement, and organisations that want an orchestrator engineers and analysts can both read

Kestra is the option to look at when the Python-centricity of Prefect and Dagster is the problem rather than the appeal. Workflows are declarative YAML, tasks can run in any language via containers, and the open-source edition is Apache 2.0 with unlimited flows and executions. It has the smallest community of the established options here, and its Enterprise Edition price is not published, which is the usual open-core trade.

Pros

  • Declarative YAML workflows are readable by people who do not write Python, which matters in organisations where analysts and platform engineers share the orchestrator
  • Apache 2.0 open-source edition with unlimited flows and executions, plus over 2,000 plugins and both code and no-code editors
  • Language-agnostic by design: task runners execute on containers or VMs, so the transformation logic can be anything
  • Event-driven scheduling is native rather than a bolt-on, which suits pipelines triggered by file arrivals or message queues
  • Kestra Cloud offers a fully managed, usage-based option for teams that want neither self-hosting nor an enterprise contract

Cons

  • Enterprise Edition pricing is not published; it is an annual per-instance subscription quoted on request
  • Nearly everything an enterprise needs is above the open-source line: SSO, LDAP, SCIM, RBAC, worker groups, audit logs, multi-tenancy, and secrets manager integration
  • Kestra Cloud pricing is described as usage-based pay-as-you-go without a published rate card, so it cannot be modelled in advance
  • Smaller community and hiring pool than Airflow, Prefect, or Dagster, which raises the handover risk
  • YAML at scale has the usual problems: large workflows become verbose and refactoring is manual
Honest Weakness: Kestra's open-source edition is real and its enterprise line is drawn exactly where a company with a security review will land. SSO, RBAC, audit logs, and multi-tenancy are all Enterprise-only, which means the free edition is fine for a team and not deployable organisation-wide at most regulated companies. Since the Enterprise price is unpublished and quoted per instance annually, you cannot find out what that costs without a sales conversation.

The Python Tax

Prefect and Dagster are both excellent and both require your orchestration layer to be Python. In a lot of organisations that is fine because the data stack is already Python. In others it forces a language into a platform team that otherwise writes Go or Java, and the orchestrator becomes the one Python service nobody wants to maintain. Kestra and Airflow both avoid the hard version of this problem: Kestra by being declarative and container-based throughout, Airflow 3.3 by adding a Task SDK for Java and Go.

Open Source Edition: free, Apache 2.0, unlimited flows and executions, 2,000-plus plugins, Postgres or MySQL backend, code and no-code editors. Enterprise Edition: not published, annual subscription on a per-instance model, adds SSO, LDAP, SCIM, RBAC, worker groups, Kafka/Redis/AMQP queues, multi-tenancy, audit logs, and secrets manager integration. Kestra Cloud: usage-based pay-as-you-go, rate card not published. As published on Kestra's pricing page, September 2026.

Visit Kestra
6

Windmill

Honorable Mention

Best for: Platform and internal-tools teams that want to turn existing scripts into scheduled jobs, approval workflows, internal apps, and HTTP endpoints without building a service for each one

Windmill is the odd one out and deliberately included, because it is what a lot of teams actually need when they think they need an orchestrator. It takes scripts in Python, TypeScript, Go, Bash, or SQL and turns them into versioned, permissioned, schedulable workflows with generated UIs. Its self-hosted tier is unusually generous. Its licensing is mixed, and the enterprise features are proprietary rather than open source.

Pros

  • Self-hosted free tier allows up to 3 workspaces, 50 users, 10 GiB of workspace object storage, and unlimited executions, which is generous for an open-core product
  • Scripts become workflows, internal apps, and HTTP endpoints from the same source, which removes a large amount of glue service work
  • Multi-language by default: Python, TypeScript, Go, Bash, and SQL, executed in isolated workers
  • Cloud Pro pricing is published and granular: from $120 a month, with developer seats at $20 and operator or external JWT users at $10 counting as half a seat
  • Worker pricing is explicit at $50 per worker per month for 2 compute units, so compute cost is separable from seat cost

Cons

  • Licensing is genuinely mixed: parts are Apache 2.0, parts are AGPLv3, and certain enterprise features are under a proprietary licence, which needs legal review in most companies
  • Free self-hosted tier excludes Kafka, NATS, and SQS triggers, audit logs, and agent workers, and restricts git sync to 2 users
  • Not a data-pipeline orchestrator: there is no asset lineage, no dataset-aware scheduling, and no equivalent of Airflow's provider ecosystem
  • Not a durable execution engine either, so it does not replace Temporal for long-lived stateful workflows
  • SSO is capped at 10 users on the free self-hosted tier and uncapped SSO requires seats on Enterprise
Honest Weakness: Windmill occupies a middle ground that is easy to buy for the wrong reason. It is excellent at replacing a pile of cron jobs, Lambda functions, and Retool screens with one versioned system. It is not a substitute for Airflow on a real data platform and not a substitute for Temporal on a real distributed workflow. Teams that adopt it as a general orchestrator find the ceiling within a year. Teams that adopt it as an internal-tooling platform tend to keep it.

The Cron Job Graveyard

Most engineering organisations have thirty to a hundred scheduled scripts running somewhere nobody can fully enumerate: a cron on a bastion host, a Lambda, a GitHub Action, a colleague's laptop. None of them is a data pipeline and none of them needs Airflow. Windmill's actual competitor is that graveyard, and the case for it is version control, permissions, observability, and a UI for the things that need one. Framed that way it is a strong purchase. Framed as an Airflow replacement it is not.

Self-hosted open source: free, maximum 3 workspaces, 50 users, 10 GiB workspace object storage, 10 SSO users, unlimited executions; excludes Kafka/NATS/SQS triggers, audit logs, and agent workers, and limits git sync to 2 users. Cloud Pro: from $120/month, developer seat $20/month, operator or external JWT user $10/month counting as half a seat, includes audit logs, distributed S3 cache, SAML. Enterprise: from $120/month, adds 24/7 priority support with 3-hour response, uncapped SSO, commercial licence. Workers: $50 per worker per month (2 compute units each); native subworkers $50/month for 8. Licensing: mixed Apache 2.0, AGPLv3, and proprietary enterprise features. As published on Windmill's pricing page and repository licence file, September 2026.

Visit Windmill

Which One Should You Pick?

Use CaseOur Recommendation
A payment, provisioning, or order flow must survive crashes and deploys and may wait days for an external eventTemporal. This is durable execution, not orchestration, and no scheduler on this page solves it. Prototype on the $150 free credit specifically to measure Action consumption, because Actions are units of workflow progress and the forecasting error is usually an order of magnitude.
Nightly batch pipelines against a dozen different databases, warehouses, and SaaS APIsApache Airflow. The provider ecosystem is the differentiator and the hiring pool is the second one. Evaluate against Airflow 3, not Airflow 2, because 3.0 GA'd in April 2025 and 3.3.0 in July 2026 added a Java and Go Task SDK.
Python-first ML and data workflows whose structure depends on data discovered at runtimePrefect. Dynamic DAGs are native rather than a workaround, and the Hobby tier is free forever for 2 users. Budget for the step change: the Team tier is $100 per user per month from 4 users.
Building a new data platform where lineage and freshness are requirements, not reportsDagster. Software-defined assets give you lineage and freshness policies intrinsically. Do not pick it to migrate an existing task-based Airflow estate; that is a rewrite, and Airflow 3's asset-driven scheduling is the cheaper answer there.
The team is not Python-first and the orchestrator must be readable by analysts as well as engineersKestra. Declarative YAML with container-based tasks in any language, Apache 2.0, unlimited flows and executions on the open-source edition. Confirm the Enterprise quote early if you need SSO, RBAC, audit logs, or multi-tenancy, because all four are above the open-source line and the price is unpublished.
Dozens of untracked cron jobs, Lambdas, and internal scripts nobody can enumerateWindmill. Versioned, permissioned scripts that become scheduled jobs, approval workflows, internal apps, and HTTP endpoints. The self-hosted free tier covers 3 workspaces and 50 users. Have legal review the mixed Apache 2.0, AGPLv3, and proprietary licensing before deploying it widely.

How we evaluated

The expensive mistake in this category is category error: buying a scheduler for a durable execution problem, or a durable execution engine for a batch pipeline. This comparison is organised so that the split is the first thing you see and the price is the second.

Each platform was assessed on the criteria in the comparison table above:

  • Which problem: data-pipeline orchestration or durable execution, stated per product rather than blurred into workflow automation.
  • Licence: the actual SPDX licence of the project, checked against the repository, including mixed-licence cases.
  • Free tier: what the free or open-source path genuinely allows, in users, workspaces, deployments, executions, or credits.
  • Published price: the vendor's own number, including the second billing dimension (credits, Actions, serverless hours, workers) that usually decides the bill.
  • Operational burden: for open-source options, the real standing cost of running it, stated as engineering commitment rather than as zero.

What we verified, and when

Last verified: September 2026. Every price, limit, and licence was read from a primary source this month:

  • Temporal pricing for the $150 / 90-day credit, the $50-to-$25 per million Actions range, Active and Retained Storage rates, and the Business support floor.
  • Astronomer Astro pricing for the managed Airflow hourly deployment, worker, dedicated cluster, and ephemeral storage rates.
  • The Apache Airflow blog for release dates and contents: 3.0.0 GA on 22 April 2025, 3.1.0 on 25 September 2025, 3.2.0 on 7 April 2026, and 3.3.0 on 6 July 2026.
  • Prefect pricing for the Hobby, Starter, Team, and Enterprise limits and serverless credit allowances.
  • Dagster pricing for the Solo and Starter prices effective 1 May 2026, the credit definition, and the serverless compute rate.
  • Kestra pricing for the open-source and Enterprise Edition feature split and the absence of a published Enterprise or Cloud rate card.
  • Windmill pricing for the self-hosted limits, cloud seat and worker rates.

Licences were verified from each project's own repository metadata rather than from marketing pages. Apache 2.0 for Airflow, Prefect, Dagster, and Kestra. MIT for the Temporal server. A mixed Apache 2.0, AGPLv3, and proprietary-enterprise arrangement for Windmill, which its LICENSE file states explicitly.

The Prefect acquisition of Dagster Labs is from the Dagster announcement and Prefect's own page, both dated 13 July 2026.

What we could not verify

Kestra does not publish an Enterprise Edition price or a Kestra Cloud rate card, so neither is comparable here and the table says so. Dagster Pro, Prefect Enterprise, and Temporal Enterprise are all quote-only. Temporal Action consumption cannot be forecast from published material for any specific workload, which is why the recommendation is to measure it on the free credit rather than to model it.

Note

Editorial independence: this is a vendor-neutral comparison with no paid placements, sponsorships, or affiliate links. Rankings reflect fit for the stated use cases, not commercial relationships. No platform here was tested hands-on; the evaluation rests on vendor documentation, published pricing, project release notes, and repository licence files, all cited above.

Frequently Asked Questions

What is the difference between workflow orchestration and durable execution?
Orchestration schedules and sequences work, usually on a timetable, usually over data: read these sources, transform, write the result, alert on failure. Airflow, Prefect, Dagster, and Kestra do this. Durable execution persists the state of a running program so it can resume exactly where it stopped after a crash, a deploy, or a multi-day wait: charge the card, wait for identity verification, compensate if the third call fails. Temporal does this. Buyers conflate them because both are marketed as workflow engines, and the mismatch is expensive in both directions. A useful test: if your workflow has a schedule, you want orchestration; if it has a trigger and a long life, you want durable execution.
Did Prefect really acquire Dagster, and what happens to Dagster?
Yes. Prefect announced an agreement to acquire Dagster Labs on 13 July 2026, uniting the two most widely adopted Airflow successors, with the combined company expected to operate under the Prefect name from August 2026. Both companies state that Dagster and Dagster+ keep their names, their open-source licence, and their roadmap, that both products remain independently supported, and that pricing for both Prefect Cloud and Dagster+ is unchanged. Dagster founder Nick Schrock and CEO Pete Hunt moved to strategic advisor roles. Those commitments are genuine and they are also the commitments every acquirer makes, so treat long-term independence as a renewal-time question rather than a settled one.
Is Airflow still the right default in 2026?
For scheduled batch data pipelines in a heterogeneous environment, yes. Airflow has the largest provider ecosystem, the largest hiring pool, and multiple managed vendors, and Airflow 3 addressed most of the architectural criticism that accumulated during the 2.x years. 3.0 reached GA on 22 April 2025 and 3.3.0 landed on 6 July 2026 with a state store for tasks and assets, a Language Task SDK for Java and Go, and pluggable retry policies. Where it is not the right default is dynamic, data-dependent workflows, where Prefect is better, and asset-centric greenfield platforms, where Dagster is better.
What does open source actually cost for these orchestrators?
For Airflow, a production deployment is a scheduler, triggerer, webserver, worker pool, and a Postgres metadata database, all of which need monitoring, capacity planning, and periodic upgrades that break DAGs. Realistically half a platform engineer as a standing commitment, plus the infrastructure. For Temporal, self-hosting means running a cluster backed by Cassandra or a supported database with real capacity planning, which is more demanding than Airflow. Prefect, Dagster, and Kestra open-source deployments are lighter but still need a database and a server. In every case the honest comparison against a managed offering is a fraction of an engineer versus the hourly or credit rate, never zero versus something.
Which workflow orchestrator has the best free tier?
For a managed product, Prefect's Hobby plan: free forever with 2 users, 1 workspace, 5 deployments, 500 serverless credit minutes a month, and 7-day run retention. Dagster+ offers only a 30-day trial. Temporal Cloud gives $150 in credits for 90 days, which is time-limited but well suited to measuring real Action consumption. For self-hosted, Kestra's Apache 2.0 edition allows unlimited flows and executions, and Windmill's self-hosted tier allows 3 workspaces and 50 users. Airflow has no vendor tier at all because there is no vendor: it is Apache 2.0 and unlimited.
Can Temporal replace Airflow for data pipelines?
Technically you can schedule work with Temporal, and in practice it is the wrong tool. Airflow and its peers ship hundreds of providers for databases, warehouses, object stores, and SaaS APIs, plus backfills, catch-up semantics, dataset-aware scheduling, and a UI built for pipeline operators. Temporal ships none of that because it is not what it is for. The reverse is equally true: Airflow cannot durably hold a workflow's state across a three-week human approval wait without you building a checkpointing system, which is precisely the thing Temporal exists to provide.
How is this different from choosing an ETL or data pipeline tool?
Orchestration is the scheduling and dependency layer. Extract and load tooling is the connector layer that moves data between systems, and transformation tooling is what turns raw tables into modelled ones. A typical stack combines all three: a managed connector product or an open-source equivalent for extract and load, a SQL transformation layer, and one of the orchestrators on this page to sequence them. Choosing a connector product is a separate decision covered in the data pipeline and ETL comparison on this site, which this page deliberately does not repeat.

About the author

is the founder and creator of LoginRadius, a customer identity platform he built and scaled to over a billion users. He is now the founder of GrackerAI, a GEO platform for B2B SaaS and cybersecurity teams, and has spent more than 15 years building identity and security products.

Related Comparisons