Brickster Newsletter

Four AI labs landed inside the same eleven-day window

Between September 1 and September 11, Unity Gateway added a hosted model from Anthropic, Google, OpenAI, and DeepSeek, all generally available. We read the release notes so you can see what actually changes when your workspace can call any of them natively.

On September 1 the release note announced that Unity Gateway now supports Anthropic’s Claude Fable 5.1 as a Databricks-hosted model, callable through Foundation Model APIs like every other model on the platform. A day later, Model Serving added Google’s Gemini 3.8 Flash the same way. Two days after that, OpenAI’s GPT-6 Astra landed, described in the release note as built “for enterprise reasoning, structured document processing, coding, agentic search, multimodal inputs, and long-context workloads.” On September 11, DeepSeek’s V4.1 Flash closed out the run, with text and image input. Four labs, four release notes, eleven days apart.

None of these are partnerships in the press-release sense. There was no joint announcement, no named customer, no dollar figure attached to any of the four. They read like release notes because that’s exactly what they are: routine additions to a gateway that has been adding models all year. The story here is the pace, not any single model’s benchmark score.

That pace matters because of what it does to the decision in front of you. Before this stretch, putting four frontier labs behind one governed endpoint meant four separate vendor integrations, four sets of credentials, four places compliance had to review. Unity Gateway collapses that into one relationship. You still choose a model per task. You no longer choose a vendor per model.

Eleven days, four labs

All four are generally available today. Not preview, not waitlisted. Turn on Foundation Model APIs and any of the four is one endpoint call away.

GPT-6 Astra is the reasoning-and-documents pick. The release note names it for enterprise reasoning, structured document processing, coding, agentic search, multimodal inputs, and long-context workloads.

DeepSeek V4.1 Flash is the newest, and the narrowest. Text and image input only, per its release note, added last of the four on September 11.

Nothing about routing is automatic. Unity Gateway makes all four callable. Deciding which task goes to which model, and what that costs, is still an architecture decision your team has to make on purpose.

So: is Databricks actually neutral, or just accumulating options? Both, for now. Four labs in eleven days is a genuine expansion of what one governed endpoint can do, and it lowers the cost of trying a cheaper model without re-plumbing your stack. It’s also exactly the kind of move a platform makes when it wants to be the default regardless of which lab wins the next benchmark round. Either read leaves you with the same task: decide, deliberately, which model earns which job, instead of defaulting to whichever one you integrated first.

Excel Add-in reaches GA

On September 10, the Databricks Excel Add-in reached general availability. It connects Excel itself to a workspace so you can browse and import governed Unity Catalog tables and views without exporting anything first, the reverse of Databricks reading .xlsx files, which has been GA since earlier this year. For every team that still ends a data request with a spreadsheet emailed back and forth, that spreadsheet can now come from the source of truth instead of a stale export.

A reference pattern for agents that survive crashes

A Databricks blog published September 8 pairs Temporal’s durable execution with Lakebase Postgres to build agents that recover from worker failures, retry safely, and keep an audit trail, demonstrated on a loan-underwriting workflow. Temporal’s Event History survives a worker restart; Lakebase stores the queryable state separately from the execution logic. The Lakebase Change Data Feed used to publish that audit trail is still in Public Preview, not GA. This is the reliability question every executive asks before trusting an agent with a real workflow, answered with a working reference implementation instead of a promise.

Marketing data gets the governed-table treatment

Lakeflow Connect expanded its marketing connectors on September 11. For acquisition: Google, Meta, TikTok, Reddit, and LinkedIn Ads. For engagement: Salesforce Marketing Cloud, SendGrid, and Marketo. For the customer relationship: Salesforce, HubSpot, and Dynamics 365. For the experience layer: Zendesk, Amplitude, and Pendo. Everything lands as governed Delta tables in Unity Catalog with no pipeline to build, and Adobe Campaign and Adobe Analytics connectors are launching next. The marketing team’s tool stack just became a data source your governance policy already covers, not a set of exports someone has to remember to reconcile.

Lakebase Postgres gets meaningfully cheaper to run

A September 10 Databricks blog walked through a Lakebase cache rework: larger shared buffers and 2MB huge memory pages to cut page-table overhead. The results, from Databricks’ own benchmarks: up to 2x throughput in one example, a cache hit rate near 100% in another, and CPU use down from 20 cores to 4 in a third. Huge pages alone cut tail latency by roughly 40% and CPU by roughly 30%, with no migration required. If you’re already running Lakebase workloads, this lands as a quiet cost reduction with nothing for you to configure.

1,500+ monthly active users

Crossed this week, alongside the Learn and Jobs launches above.

Quick links this week:

→ brickster.ai/exam-prep: Practice sets for the Databricks certification exams, unlimited runs, no signup.

→ brickster.ai/jobs: This week’s Databricks openings, refreshed daily.

→ brickster.ai/digest/week: Everything we tracked this week, one page.

🧩 Not sure which model to route a task to?

The Brickster Assistant searches our full archive and answers with citations.

Enjoying the Brickster Newsletter?

Forward this to a colleague who works with Databricks.