01Databricks put an agent on your laptop. Genie Code CLI works on the files in front of you, under your own identity. The surprise is on the invoice: it isn’t billed as Genie.
02Four things for the calendar. Four models switch off by November 5. Genie One learns to run the numbers. One CLI upgrade can stop a deployment. And a budget alert nobody owns.
03Three Genies, one name. Our explainer on which Genie does what, and what each one costs.
01
💻 Databricks put an agent on your laptop
01 / THE LEAD
Databricks put an agent on your laptop. The meter runs somewhere else.
Genie Code CLI went into Beta on October 6. It works on the files on your machine, under your own Databricks identity. And it’s billed through Unity Gateway, not as Genie, which is the detail most teams will miss until the invoice arrives.
Until last week, Databricks’ coding agent lived in a browser tab, a long way from the repository on a developer’s machine. Genie Code CLI closes that gap. It runs in the terminal and reads and edits the project files already open. It runs commands, queries your data, and builds Databricks pipelines and apps. It acts as the developer, under their Databricks identity, so it can reach exactly what they can reach. Nothing more.
You don’t have to switch tools to get most of that. Databricks’ ug tool points an agent you already use, Claude Code or Codex among them, at Unity Gateway, Databricks’ control point for AI models. You can call Databricks-provided models or bring your own model provider account. Keep the agent your developers like. Change where the models come from, and who sees the bill.
The bill is the part worth reading twice. The CLI’s docs are blunt: “It is not billed under Genie pricing,” and it’s “not impacted by Genie budgets.” You pay for model usage through Unity Gateway instead. The control that does apply is a Unity Gateway budget: a monthly limit per person that can send an alert or block further requests. Blocking works from an estimate of spending, and requests already running are allowed to finish. Treat the limit as a brake, not a wall.
Starting a trial is an afternoon of admin, not a project. An admin confirms the workspace sits in a supported Unity Gateway region with Unity Catalog, Databricks’ governance layer, turned on. The developer needs Databricks CLI 1.0.0 or later, installed separately. Then they install Genie Code CLI and start it inside the project folder. The requirements page has the detail.
Then the catches. During Beta, Databricks picks the model, currently GPT 5.6, and you can’t choose another. The CLI doesn’t yet match the browser version feature for feature. The two don’t share skills or instructions either, so whatever your team taught one, it will have to teach the other. And Databricks classes Beta features as not for production, with no service-level guarantee.
“It is not billed under Genie pricing.” Databricks docs, Genie Code CLI.
So which one? If your developers already live in Claude Code or Codex, connect them to the gateway first. It’s the smaller change, and they keep the tool they know. If you want an agent that arrives already fluent in Unity Catalog, pipelines and apps, give Genie Code CLI the same task and watch both. One developer, one week, one job outside production, with the spend in front of you. The invoice will tell you more than the launch note.
Your move this week
Business leaders: name an owner for the trial, decide what a useful result looks like, and agree how much it can spend.
Data leads: give one developer one week on a task outside production. Set an early alert and a higher block. Review the work and the total cost before anyone else joins.
Each of these will land on somebody’s desk this month. Better it lands there on purpose.
AI MODELS
Four models switch off, and the obvious replacement costs more
Databricks is retiring four models from pay-per-token serving, where you pay for the text processed. DeepSeek V4 Pro (0813), Kimi K2.7 and Inkling go on October 30. DeepSeek V4 Flash (0731) follows on November 5. Databricks names the replacements: DeepSeek V4.1 Flash for both DeepSeek models, Kimi K3 for Kimi, and GLM 5.3 or Kimi K3 for Inkling. The job is finding which apps and scheduled jobs still call the old ones, and testing the swap before the date, not after it. Check the price while you’re there: at published standard rates, V4.1 Flash costs about twice as much for input and more than four times as much for output as the V4 Flash it replaces.
Genie One will now do the maths on your spreadsheet
Genie One, Databricks’ chat app for business users, can now write and run code to forecast a number or test a what-if, including on an uploaded CSV or Excel file. Code execution is in Public Preview, and it runs as you, so it can only touch the data you already can. The upload itself is a separate Beta that a workspace admin turns on, and each CSV or Excel file must be under 200 MB. The analysis runs on serverless compute and is billed as its own sandbox usage. For a first outing, hand it a spreadsheet whose answers you already know, and check its working.
Databricks CLI 1.20.0, released October 7, removes the Terraform engine behind bundles, the packages teams use to ship pipelines and jobs. Older bundles still migrate to the direct engine automatically. The difference is what happens when that fails: the deployment now stops. Teams that still need Terraform have to stay on 1.19.x. One quieter change deserves a look too. Delete a field from the deployment file, and the next deployment resets it to the default. Rehearse the upgrade in a test environment before it reaches production.Our September article has the background.
Databricks has published a guide to keeping SQL warehouse costs in check, the compute behind your dashboards and queries. The advice is sound: tag spending by team, set budgets, add alerts, review usage. Two limits decide how far it gets you. These budgets don’t stop a warehouse running, and the alert email can arrive up to 24 hours later. That’s the opposite of the Unity Gateway budgets in the lead story, which can block a request outright. So put a name next to every alert, and agree the first move in advance, whether that’s slowing a dashboard’s refresh or resizing the warehouse.
Three Genies, one name. Which does your team need?
Ask three people what “Genie” means and you’ll get three answers. Our explainer sorts it out. Genie One is the chat app for business users. Genie Agents answer questions about your company’s data. Genie Code helps developers write and fix code. Pricing for each is in there too. It covers the browser tools; the CLI in this week’s lead is newer, and bills on its own.
