Databricks Genie Code: what it does and what it costs

Databricks Genie Code is the coding agent built into the Databricks workspace. You describe a job, such as "load yesterday's orders, drop the test accounts and chart revenue by region", and it plans the steps, finds the tables, writes the code, runs it, reads the output and fixes what failed. It works in notebooks, the SQL editor, the Lakeflow Pipelines Editor, AI/BI dashboards and MLflow.

Since 8 July 2026 it is also the one Genie product you pay for: each person gets 150 DBUs of use free a month, and the rest is billed at about $0.07 a DBU in US East.

What Genie Code does

Genie Code is one of the 3 Genies, the one for people who build. Genie vs Genie One vs Genie Code shows how it fits with the other 2. What sets it apart from a chat assistant is that it acts:

• It runs what it writes. It runs a cell, reads the result or the error, changes the code and runs it again, the same loop an engineer would follow.
• It knows your catalog. It works with Unity Catalog tables, columns and lineage, so it can find the orders table and see what feeds it.
• It builds whole things. Pipelines in the Lakeflow editor, dashboards, MLflow experiments and metric views, where the metric view editor has it built in.
• It can be taught. You give it instructions for the workspace and for yourself, reusable skills, memory across sessions and MCPModel Context ProtocolAn open standard for connecting AI applications to tools and data. A system offered once as an MCP server works in any application that speaks MCP. servers for outside tools. A full-page view runs several chats side by side.

If you only want an explanation, say so: Databricks' own example is "Just explain this to me, don't run any code."

What it is allowed to do

An AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. that runs code in your workspace needs limits, and Genie Code has 3:

1. Your permissions, not more. It can only read and change what your Unity Catalog grants allow.
2. An approval step. By default an AI classifier approves each tool action and is designed to block actions that go beyond your request. You can switch to approving every action yourself, which is the safer choice for anything that writes to production tables.
3. What it sends. To do a task it sends your prompt, table and column names and descriptions, your code, sample rows and cell outputs to the model, within your permissions. Its agent mode needs the partner-powered AI features switched on for the account and workspace, and a region where Genie Code is offered. Without them it still answers, from a Databricks-hosted model, but does not act on its own.

The same rules apply to any coding agent. How to add human approval to an AI agent explains where the gate belongs.

What Genie Code costs

The free part covers Genie's model usage only. When Genie Code runs a query on a serverless SQL warehouse, that warehouse bills you as it would for a person, which our Databricks cost calculator prices.

Databricks does not publish how many DBUs a typical task uses, so the only honest way to forecast is to watch a month of your own use. Try your team's shape:

Try itWhat will Genie Code cost your team?

People using Genie Code

20people

10100

DBUs a month, light user

80DBUs

150 free

DBUs a month, heavy user

600DBUs

150 free

Heavy users

10%25%50%

Per-user budget that blocks

None$20$50

Group	People	Free DBUs each	Billed DBUs each	A month
Light users	15	80	0	$0.00
Heavy users	5	150	450	$157.50
Genie Code, a month				$157.50

See the bill in system tables

Genie usage is in system.billing.usage like everything else. Filter on the product, because the billed SKU, Serverless Real-Time Inference, is shared with other services:

SELECT
  identity_metadata.run_as        AS person,
  usage_metadata.genie.surface    AS genie,
  sku_name,
  SUM(usage_quantity)             AS dbus
FROM system.billing.usage
WHERE billing_origin_product = 'GENIE'
  AND usage_date >= date_trunc('MONTH', current_date())
GROUP BY ALL
ORDER BY dbus DESC;

Rows with sku_name = 'GENIE_FREE_USAGE' are the free part. They have no list price, so a join to system.billing.list_prices leaves them out, and they only exist from 20 July 2026. The genie column says GENIE_CODE, GENIE_ONE or GENIE_AGENTS.

Put a cap on it

A Genie budget does more than send email: it can stop spending. An account admin creates a budget in the account console with the Unity Gateway resource type and the tag databricks-product: genie, and nothing else. Then:

• A shared threshold is one monthly pool for everyone in scope. With Block usage checked, everyone stops when it runs out.
• A per-user threshold stops one person at their own limit, with overrides for groups who need more.

Databricks recommends alerts only on the shared pool and blocking per user, so one heavy user cannot stop the whole team. Budgets cover the paid part only; nobody loses their 150 free DBUs.

Is it any good?

Databricks says that on real-world data science tasks Genie Code more than doubled the success rate of leading coding agents, from 32.1% to 77.1%. It has not said which tasks, which agents or how success was judged, so treat the figure as a claim to test. Its edge is context: it already knows your tables, permissions and lineage, which a general agent such as Claude Code has to be given. A fair trial is a week of real notebook work with each, counting the fixes you had to make by hand.