If you run Databricks, you can switch on the new Genie and Genie Ontology today, but the always-on background agent that headlined the launch has no availability date and no price of its own. Databricks announced the second-generation Genie on November 11, 2026 and lists Genie and Genie Ontology as generally available, while Genie Search, Ontology Manager and the always-on mode are "coming soon" to preview and beta programs. Pilot the generally available half now, while user usage of Genie One and Genie Agents is still free through January 31, 2027, and write your spend and permission rules before the background agent reaches your workspace.
The timing matters because the free period distorts the numbers. Whatever adoption you measure in December and January is usage you did not pay for, and the bill starts on February 1.
What Databricks Shipped on November 11, and What It Only Announced
Two of the five pieces in the launch are generally available. The rest are announced for preview or beta with no date, per Databricks' own release:
| Piece | Status on November 11, 2026 | What it does, per Databricks |
|---|---|---|
| Genie | Generally available | The conversational agent, now reachable from browser, mobile, email and tools including ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot |
| Genie Ontology | Generally available | Captures definitions, metrics, people and processes, learning from dashboards, queries, notebooks, external BI tools and an org-structure connector |
| Genie Search | Coming soon, preview or beta | Search across the lakehouse and apps like Google Drive, Jira and Confluence through 100+ connectors, indexed in place, ingested via Lakeflow Connect, or queried live over MCP |
| Ontology Manager | Coming soon, preview or beta | Governs business context and flags frequently asked concepts that need refinement |
| Always-on Genie | Coming soon, preview or beta | Runs in the background on its own compute, delegates to subagents, chases missing inputs and asks for decisions through a priority inbox and email |
Independent coverage agrees on the split. ITBrief and MLQ both report Genie and Genie Ontology as GA and the always-on capabilities as preview. One third-party setup guide on DEV Community lists Genie Search as GA. The vendor release does not, so treat the release as the record until your account team confirms the stage in writing.
The always-on mode is the part with no precedent in your current Genie rollout. Databricks describes a user giving it a goal and boundaries, after which it tracks progress across multi-step work, follows up on missing inputs and delivers each new alert as its own unread email. The examples in the release are meeting prep, quarterly forecast tracking, customer usage monitoring and a weekly executive revenue brief. Every one of those runs while nobody is watching the session.
The customer evidence is thin and vendor-supplied. HelloFresh's associate director of data science and engineering says Genie "reduced the time to identify ad winners on Meta by 75%," and CEO Ali Ghodsi says Genie "has made me a 10x more productive CEO." Both quotes come from the Databricks release, and the 10x figure is a personal claim rather than a measured result.
How Genie Is Billed Once the Free Period Ends
Genie bills on usage, metered in DBUs, with a monthly free allowance per user and a promotion that expires on January 31, 2027. The release FAQ says customers pay for what they use rather than per seat, and gives no dollar figures. The billing mechanics sit in the docs.
Databricks' Genie cost documentation (last updated September 11, 2026) says each user "receives 150 Databricks units (DBUs) per month free," resetting on the first of each month. Service principals get no allowance. Usage above it bills under the Serverless Real-Time Inference SKU, whose name varies by cloud and region, and every Genie row in system.billing.usage carries billing_origin_product = 'GENIE', which is how you separate it from other products sharing that SKU.
Two promotions run until January 31, 2027. User usage of Genie One and Genie Agents is free, and a 25% discount applies to all billed Genie usage, applied inside DBU metering so usage_quantity already reflects it. Service principal usage is billed throughout. The same page tells you to multiply by pricing.effective_list.default from system.billing.list_prices to get list cost.
That gives you a February problem. The free period and the discount end on the same date, so a team that pilots through January sees a cost of roughly zero, then a bill with neither promotion applied (only the 150 DBU monthly allowance remains, and your negotiated rate sets the price). Price your pilot at list from day one using that system table, or the business case will be wrong by the whole bill.
The allowance has also been described in two different units. At the Genie One launch in June 2026, Databricks said organizations "get up to $10 free for every user every month." The current docs say 150 DBUs. Ask your account team which governs your contract and what 150 DBUs is worth at your negotiated rate.
The always-on agent has no published unit at all. Databricks says it runs on "its own compute" and has not said whether that compute bills as Genie inference, as serverless compute, or as something else.
Unity Gateway Budgets Cap Model Spend Approximately and Exclude Compute
The release says admins can use Unity Gateway to choose which tools Genie may call, set approval rules and set budgets or spending caps by team or user. The documentation is more careful about what those caps do.
Microsoft's Azure Databricks guide to Genie budgets, updated November 11, 2026, says to create a budget with the Unity Gateway resource type and the databricks-product: genie tag, and warns that adding any other tag stops the budget from tracking Genie at all. A budget can combine a shared pool with per-user thresholds and group overrides. Databricks recommends blocking on per-user thresholds rather than on the shared pool, because a blocked shared pool stops Genie for everyone in scope.
The limitations section is the part to read before you promise finance a ceiling:
- "Budgets track LLM usage. Compute used to run queries, such as a SQL warehouse, is billed separately and is not included in Genie budgets."
- Block thresholds "are enforced approximately," and "active Genie requests in progress when the threshold is reached are not interrupted."
- "Do not use this feature as a way to ensure an absolute spend cap on final billed amounts."
- Each budget supports at most four shared thresholds and 100 per-user overrides.
The general Unity Gateway budgets page carries the same warning, and notes that users can be blocked before they reach a threshold or overshoot it before enforcement lands. Its Genie cost FAQ uses a $30 per-user monthly threshold as its worked example.
For a person typing questions, approximate enforcement is tolerable. For a background agent that runs long, multi-step jobs, it is a different exposure: the uninterrupted in-flight request is the agent's normal working state, and the warehouse queries it fires sit outside the budget entirely. We covered related cap questions in Salesforce's Agent Fabric spend caps and in AWS Kiro bots drawing on interactive credits. For an agent, budget the model spend and the query compute separately.
Where the Always-On Agent Changes Your Permission Model
Genie follows existing Unity Catalog permissions, so it sees only the data each person may access. That is the right default for an assistant answering one user's question, and it is what Databricks commits to in the release.
A background agent raises questions the release does not answer. Whose identity does a long-running Genie job run under after the requesting user's session ends? If it runs as a service principal, the docs say service principals get no free allowance and are billed for all usage. Databricks has not said which it is.
The output path also leaves the lakehouse. Each always-on alert arrives as its own email, with updates added to the same thread. Whatever that email contains is governed by your mail system's retention and forwarding rules from then on, outside Unity Catalog's. Ask whether alert emails carry data or only a link, before a weekly revenue brief starts landing in forwarded threads.
Then there is what the ontology learns from. Genie Ontology builds definitions from your dashboards, queries and notebooks, which means it inherits whatever conflicts already exist between them. Our semantic layer comparison and Grab's five months of analytics agents both found that governed metric definitions, more than the model, set the accuracy ceiling. Headspace's data engineering director says the same in the Databricks release: "AI is only as good as the definitions behind it."
How This Compares With Snowflake's Agent Billing
Snowflake bills the same way in outline, and that is the useful comparison if you run both platforms. Snowflake's AI pricing docs say "Cortex Agents are billed in AI Credits per million tokens processed," at $2.00 per AI Credit under global routing and $2.20 under regional, while warehouses stay on Platform Credits. The page also warns that agent costs "are additive across the underlying services the agent invokes."
Both vendors split model spend from query compute, and both caps cover only the model half. Snowflake took its own Cortex AI Gateway GA with per-user quotas and previewed a business ontology earlier this month. Snowflake's scheduled agents shipped earlier too, and in August we found they ran as all of a user's roles by default, which is the class of problem to test for when Genie's background jobs arrive.
If you are comparing Genie with Snowflake Cortex AI or ThoughtSpot for self-service analytics, our text-to-SQL comparison still holds on accuracy. What has changed is that the incumbent Databricks product now does scheduled, unattended work, and it is free to try for another eleven weeks.
The competitive pressure is visible in the calendar. MLQ notes the Genie launch came a day after Anthropic released Claude Sheets and Meetings for the same office users, and Databricks has made Genie callable from inside ChatGPT, Claude and Copilot Cowork.
What to Do Before January 31, 2027
This Week:
- Create a Genie budget with the Unity Gateway resource type and only the
databricks-product: genietag. Set a per-user threshold with Block usage, and set the shared pool to email only, as Databricks recommends. - Run a query against
system.billing.usagefiltered onbilling_origin_product = 'GENIE', joined tosystem.billing.list_prices, and put the list-price cost of your current usage on one page. That is your February number. - Ask your account team three questions in writing: when the always-on mode enters preview, what compute it bills against, and whether the $10 or the 150 DBU allowance governs your contract.
This Month:
- Pick one team with certified metric views and turn on Genie Ontology there first. Review what it learned from your dashboards before you widen it.
- Put a monthly cap on the SQL warehouses Genie queries hit, because the Genie budget does not count them.
- Write the approval rules in Unity Gateway for any tool that sends, writes or opens a ticket. Start with read-and-report jobs only.
Before January 31, 2027:
- Decide which users keep Genie access after the free period ends, using the list-price figure from step two rather than the December dashboard.
- If the always-on preview has opened, test it as a low-privilege user and record which identity its background jobs run under, what its alert emails contain, and where that compute appears in your bill.
The Bottom Line
Databricks has shipped the conversational half of Genie and announced the autonomous half. The generally available half is cheap to evaluate until January 31, and your existing Unity Catalog permissions govern it, which makes it a low-risk pilot. The always-on agent is a different purchase: it acts without a person in the session, its compute is unpriced, and the budget tool Databricks points to excludes compute and says in its own docs that it is not a hard cap. Treat it the way you would treat any consumption-billed warehouse feature: caps and identity first, broad rollout second.
Run the list-price query this week and send the three questions to your account team before the preview opens.
Continue Reading
- Snowflake Copilot vs Genie vs ThoughtSpot: Copilot Is the Wrong Tool
- Choose Native Semantic Layers for One Warehouse, Cube for More
- Snowflake Takes Cortex AI Gateway GA, Previews Business Ontology
- Snowflake CoCo Automations Run on Schedules Without Tool Prompts
- Test Agent Fabric Spend Caps Before Retiring Your AI Gateway
- Grab Ran Analytics Agents 5 Months. Context Was the Ceiling.