Sheila S.

Images and bounding boxes in prompt_jev(): prompt_jev() now supports image classification and scoring with PNG/JPEG inputs. New box questions identify bounding boxes matching a prompt, while multiple_choice picks every label that applies instead of just one.

GPT-6 Luna in prompt(): Use model := 'gpt-6-luna' in prompt() to access OpenAI's GPT-6 Luna. reasoning_effort now defaults to 'none' for faster, lower-cost responses; set it to 'low' or higher for more complex tasks.

Create, edit, share, and delete Dives with Agent Chat: Agent Chat can now create and manage Dives directly. Ask it to build a Dive from your data, add charts, update filters or queries, edit titles and descriptions, share underlying data, or delete a Dive. Changes open automatically for review, and edits are preserved in version history. Deletions are permanent.

In-region inference for Agent Chat: Agent Chat now runs model inference within the same geography as your selected cloud region (US or Europe).

DuckDB 1.5.6 support: MotherDuck supports DuckDB 1.5.6. Refer to the changelog and the official announcement for details.

Note: Our install scripts ensure you always have a MotherDuck-supported DuckDB client version.

Lower prices for AI functions: The prompt() and embedding() functions now cost 50–60% less on every OpenAI model, which brings them in line with other providers. Each AI Unit covers more input and output tokens. Refer to AI function pricing for the new rates.

Billing starts for prompt_jev(): Starting October 1st, prompt_jev() usage is billed in AI Units. Input tokens are metered, while output tokens are free. Usage counts toward the same daily AI functions limits as prompt() and embedding().

2x faster prompt_jev(): prompt_jev() now sends twice as many requests in parallel, so classifying a large table takes about half as long as it did last week.

Select a range of cells in query results: Drag across the data grid to select multiple cells, or hold Shift and click to extend a selection. Hold a drag near the grid's edge to scroll to offscreen cells. You can copy the selected range as a rectangular block.

Descriptions for service accounts and access tokens: Users can add an optional description when creating a service account or access token to record what it's for or who owns it. Searchable descriptions appear under the name wherever service accounts and tokens are listed.

Flight sizes: Flights now run on one of three sizes: F4 (0.5 vCPU, 4 GB), F16 (2 vCPU, 16 GB), or F32 (4 vCPU, 32 GB). F16 is the default on every plan except Lite (with limits), which runs on F4. F32 is available on Business and Enterprise. Set a Flight's size with instance_type on MD_CREATE_FLIGHT or MD_UPDATE_FLIGHT. Each run is billed at per-size hourly rates.

Admins can cancel any Flight run: Users with the Admin role can cancel a pending or running run of any Flight in their organization, including Flights they don't own, using SQL, MCP, or the CLI. Editing and deleting a Flight is still limited to its creator.

Pasting code into the Flights editor: Pasting Python or requirements.txt from Slack or a web page into the Flights editor no longer fails with SyntaxError: invalid non-printable character U+00A0. Non-breaking spaces are converted to regular spaces on paste.

Service Accounts page loads for long usernames: The Service Accounts page in Settings now loads for organizations with many or long service account usernames.

Cross-region DuckLake data paths: Users can now create a DuckLake on external object storage, including in regions outside your MotherDuck organization.

Lower memory use for LIMIT queries on wide tables: SELECT * FROM <table> LIMIT n, LIMIT ... OFFSET, small ORDER BY ... LIMIT, and SAMPLE now scan only the row groups holding the requested rows instead of every row group in the table.

Classify text with prompt_jev: prompt_jev sorts text (e.g., support tickets, reviews) into categories directly in SQL using Jev, a TypeSafe AI model that returns a true/false probability, a label, or a score instead of writing prose for 50x speed improvements at 1% of the cost of a frontier model. Ask one question or several in a single call about a piece of text: whether a statement is true, which label from a list fits, or where the text sits on a scale, and MotherDuck batches rows for you in response.

Note: prompt_jev is currently available on AWS in us-east-1, us-west-2, ap-northeast-1, and ap-southeast-2. Availability in additional cloud regions is coming soon.

User agent details in the Duckling Overview UI: The Duckling Overview UI now includes a User Agent column to show Admins the associated client that issued each query, including any custom user agents you have set, for easier sorting, filtering, and search.

Terraform provider: MotherDuck's official Terraform provider supports Terraform 1.5 and later, enabling you to manage databases, shares, roles, secrets, and service accounts as code.

AWS Marketplace: MotherDuck is now available on AWS Marketplace via Private Offer, enabling you to consolidate billing and transact directly through AWS. Get in touch to learn more.