dbt Summit 2026: What Changed for Data Teams, and What Didn’t

Summary

• dbt v2 is generally available. The Rust-based engine once called Fusion is now just “dbt”. dbt Core carries on as dbt v1, still under Apache 2.0.

• dbt State is generally available. It skips models whose results wouldn’t change. dbt Labs says early adopters averaged 15–30% compute savings.

• Lake Compute (private beta) runs selected models on a DuckDB-based engine, directly against Apache Iceberg tables.

• dbt Wizard, dbt Charts and the Fivetran Context Layer are all in preview or beta: an agent grounded in your dbt project, dashboards as code, and unstructured context for agents.

• Our take: tests, contracts and semantic definitions decide whether any of this pays off. That work hasn’t changed.

In mid-September, dbt Summit 2026, the conference formerly called Coalesce, brought more than 2,000 data professionals to The Cosmopolitan in Las Vegas, according to the organizers. It was the first one since Fivetran and dbt Labs completed their merger on June 1, 2026. Fivetran + dbt Labs framed the week around one thesis: the foundation that makes analytics trustworthy is the same foundation that makes AI trustworthy.

I was in Las Vegas for it. At Cheesecake Labs we build data platforms and production AI systems for US companies, and dbt is part of most of the stacks we work in. So I went with one question: what changes for the teams we work with, and what stays the same?

This article covers what was announced, what is generally available versus still in beta, and how we’d act on it.

What was announced at dbt Summit 2026?

Here’s what was announced and where each piece stands.

Announcement	Status (Sept 2026)	What it does
dbt v2	GA	The Rust-based engine formerly called Fusion, now the default dbt
dbt State	GA	Rebuilds what changed and skips the rest
Anthropic integration and ChatGPT plugin	GA	Bring structured dbt project context into those AI tools
dbt Wizard (dbt platform)	Public Preview	An agent for analytics engineering, grounded in your dbt project
Wizard Explore Mode	Public Preview	Lets people ask questions about their data in plain language
Wizard CLI	Public Beta	The same agent, in the terminal
Wizard Desktop	Private Beta	A desktop app for longer, parallel work
dbt Charts	Public Beta	Dashboards as code: version-controlled YAML that lives next to your models
Lake Compute	Private Beta	A DuckDB-based engine that runs selected models on Iceberg tables
Fivetran Context Layer	Private Beta	Combines structured data with unstructured sources, like docs and Slack threads, as context for agents

Source: dbt Labs, “Everything we announced at dbt Summit and why it matters”.

The announcements build on two products that are already GA: the dbt Semantic Layer, which governs metric definitions, and the dbt MCP Server, which exposes models, metrics, lineage and test results to agents.

What does dbt v2 mean for teams on dbt Core?

It’s the first major version in almost five years, as the keynote noted, and the naming changed:

• dbt (formerly Fusion) is the full v2 distribution. It’s free to use and adds local features that dbt OSS doesn’t have. Its license is not Apache 2.0.

• dbt OSS (formerly dbt Core v2) is the subset of v2 made only of Apache 2.0 components.

• dbt v1 is dbt Core as you know it, still Apache 2.0. dbt Labs says 1.13, once released, will be the final 1.x minor version, with security and installation patches expected for 3 to 5 years.

The headline is speed and feedback. According to dbt Labs, v2 parses a 10,000-model project up to 10x faster than v1, and it surfaces errors, column checks and lineage before anything runs. Faster feedback helps the engineer at the keyboard, and it also helps an agent that writes and checks models.

Adapter coverage decides your timing. BigQuery, Databricks, DuckDB, Redshift and Snowflake are GA. ClickHouse and Spark are in beta, and Athena, Fabric and Postgres are coming soon (dbt Labs announcement post). One detail to watch: pip install dbt-core now installs dbt OSS, so pin your versions if you plan to stay on v1 for now (What’s the difference between dbt and dbt OSS?).

Our take: In March we wrote that dbt Core was enough for most engagements. For v1, that is still true. If you run on one of the five GA platforms, start a v2 migration plan now. If you’re on Postgres, Fabric or Athena, wait. In both cases, bring in legal early: a team that approved dbt Core under Apache 2.0 should review the full dbt license before it reaches production, or standardize on dbt OSS.

Is dbt State worth turning on?

Probably, but measure it first.

dbt State reads model SQL and warehouse metadata to work out whether a model’s results would actually change. It then builds, skips, clones or defers each node, with no selection syntax or manifest scripts to maintain. It also lets teams set freshness at the model level. It became GA on Sept 16, 2026, and runs wherever you run dbt: the dbt platform, Airflow, Dagster, GitHub Actions or a laptop. It supports Snowflake, BigQuery, Databricks and Redshift (dbt State is GA).

The numbers dbt Labs shared are customer-reported:

• Early adopters averaged 15–30% compute savings.

• Virgin Media O2 reported 25% savings on both job run time and BigQuery compute costs.

• RxBenefits reported a 59% cut in warehouse costs on scheduled jobs.

dbt State is a paid, usage-based feature, and you don’t need v2 to try it: it works with both v2 distributions, and with dbt Core v1.7 or later through a separate plugin. It’s billed on daily active target tables, meaning each model or test skipped, cloned or reused in a day, and it comes with a 30-day free trial.

The savings depend on how much of your DAG actually changes between runs. A project that rebuilds everything every hour on mostly static sources stands to gain the most. Run the trial against at least a week of real schedules, compare warehouse spend against the State bill, and decide from that.

Where does Lake Compute fit?

Lake Compute is a single-node SQL engine built on DuckDB. It runs dbt transformations directly against Apache Iceberg tables. You tag the models you want to run there, one or a hundred, and the rest stay on your warehouse. It can work alongside Fivetran’s Managed Data Lake Service, which was already GA and keeps data as Iceberg tables in the customer’s own cloud storage.

Our take: dbt Labs pitches Lake Compute as a way to right-size compute, not to replace the warehouse. Our read is that a single-node engine suits small and mid-sized models best, and we’d test it before moving heavy joins there. It’s in private beta, so we’d pilot it on non-critical models and keep it out of this quarter’s roadmap.

What does “context engineering” mean for analytics engineers?

In its announcement post, dbt Labs puts it this way: “context engineering is analytics engineering with a new last mile.” The argument is that the people who already define models, tests and metrics are the best placed to give agents reliable context.

The pieces fit together like this:

• The Semantic Layer holds governed metric definitions.

• The MCP Server and the new Anthropic and ChatGPT integrations bring that context into the AI tools people already use.

• dbt Wizard is an agent grounded in your dbt project, available in the platform and the terminal, with a desktop app in private beta. Like dbt State, it’s a usage-based paid feature.

• The Fivetran Context Layer adds unstructured knowledge, like docs and Slack threads, and stores it in your warehouse in an open format any agent can use.

The keynote slide for the Context Layer split the flow into three steps: meaning, discovery and reach. Business, operational and analytical sources are parsed, indexed and cataloged. That knowledge lands in your warehouse in Agents Schema, an open, customer-owned format. From there it reaches agents such as Claude Code or Gemini CLI through an MCP connection.

Our take: This matches what we see in client work. The failure is rarely the model. It’s an agent reading undefined metrics and answering with confidence. Teams with tested models, documented columns and a semantic layer will get value from Wizard and the MCP Server quickly. Teams without them won’t.

What didn’t change?

• The fundamentals still decide the outcome. Tests, contracts, documentation and metric definitions determine whether agents give correct answers.

• Most of the AI surface is in preview or beta. Plan production work around the GA pieces: dbt v2, dbt State, the Semantic Layer, the MCP Server and the Anthropic and ChatGPT integrations.

• Open source stays. dbt v1 and dbt OSS are both Apache 2.0.

• dbt still works with any ingestion tool. The merger tightens the link between Fivetran and dbt. Teams that load data with other ingestion tools or custom connectors still run dbt the same way.

What we’d do in the next 90 days

• List every dbt project with its version and adapter, and plan the v1-to-v2 path where the adapter is GA.

• Run the dbt State trial on real schedules (no v2 migration needed on v1.7 or later) and compare warehouse spend before and after.

• Close the gaps in tests, documentation and semantic definitions before pointing any agent at the project.

• Pilot the MCP Server or dbt Wizard on one well-governed domain, with a small group of users, and budget for usage.

• Keep Lake Compute and the Fivetran Context Layer on the watch list until they reach GA.

How Cheesecake Labs can help

We’re a senior engineering partner that modernizes systems and data, then takes AI from assessment to scale. We build data platforms with dbt, often alongside Snowflake and Databricks.

If you’re weighing a dbt v2 migration, or you want agents working on data you can trust, let’s talk. We’ll tell you exactly what we’d build, and what we wouldn’t.

FAQ

When and where was dbt Summit 2026? Sept 15–18, 2026, at The Cosmopolitan in Las Vegas. Partner Day was on Sept 15, and the two keynotes were on Sept 16 and 17. The event was previously called Coalesce.

Is dbt Core going away? No. It’s now called dbt v1 and stays open source under Apache 2.0. dbt Labs says 1.13, once released, will be the final 1.x minor version, with security and installation patches expected for 3 to 5 years.

What’s the difference between dbt and dbt OSS? Both are dbt v2 and use the Rust-based engine. dbt is the full distribution: free to use, not Apache 2.0, and it adds local features and seat-based paid options that dbt OSS doesn’t have. dbt OSS includes only the Apache 2.0 components. Usage-based features like dbt State and dbt Wizard work with both.

Does dbt State cost extra? Yes. It’s usage-based, billed on daily active target tables, and it comes with a 30-day free trial.

Can I watch the sessions? Yes. dbt Labs has published the keynotes and breakout sessions on demand, linked from the dbt Summit page at getdbt.com/dbt-summit. The full list of announcements is public in the dbt Labs announcement post.