Databricks Summit Opens and Anthropic Claude-4 Cutover Lands

<p><strong>Anthropic's June 15 hard cutover is live</strong> — pinned <code>claude-4</code> model aliases now hard-fail across the API, Bedrock, and Vertex AI, while the Agent SDK billing split activates per-user credit pools simultaneously. The <strong>Databricks Data+AI Summit</strong> opens the same day with MLflow 3 GA bringing agent trace capture and...

Highlights

  • Databricks Data+AI Summit opens in San Francisco with MLflow 3 taking center stage: full agent trace capture, multi-turn evaluation harnesses, and a unified mlflow.agents namespace now GA (Databricks)
  • Lakebase serverless Postgres is production-ready on AWS as of today, giving RAG and data-mesh teams sub-second SQL over lakehouse data without instance management (Databricks)
  • Anthropic's June 15 hard cutover is live — pinned claude-4 model aliases return errors across the direct API, Amazon Bedrock, and Vertex AI as of today; migration is not optional (Anthropic)
  • Agent SDK per-user credit pool split is simultaneously active today; operators sharing a single pool who haven't updated billing attribution logic will route charges incorrectly from this point forward (Anthropic)

Key Signals

  1. June 15, 2026 · Data+AI Summit Day 1 keynote MLflow 3 GA reframes agent observability as an evaluation primitive

    MLflow 3 promotes LLM calls, tool invocations, and memory reads to first-class logged entities via the new mlflow.agents span model. Teams can now close the loop between production traces and benchmark regression gates without bespoke instrumentation — and because traces are stored as Delta tables, Unity Catalog lineage and access controls apply automatically. (Databricks)

  2. June 15, 2026 · Anthropic API and Agent SDK Anthropic cutover forces parallel triage on two breaking changes

    As of today claude-4-sonnet, claude-4-opus, and related pinned identifiers are hard-failed; valid targets are claude-sonnet-4-5, claude-opus-4-8, and claude-fable-5. Simultaneously, the Agent SDK billing split activates per-user credit pools — meaning teams face two independent breaking changes in the same deploy window. (Anthropic)

  3. June 15, 2026 · Data+AI Summit Databricks Mosaic AI adds agent lineage inside Unity Catalog governance

    Mosaic AI's compound AI orchestration layer logs every retriever call, agent handoff, and judge evaluation as a versioned Unity Catalog artifact. Agentic pipelines built on Databricks get auditability and rollback at the data platform level rather than requiring a separate observability sidecar — a direct competitive answer to LangGraph's stateful execution model for teams already on the lakehouse. (Databricks)

Why It Matters / What To Watch

  1. Triage both Anthropic breaking changes before monitoring metrics
    • Grep for claude-4 strings in environment variables, LiteLLM proxy configs, and LangChain model wrappers — not just application code — before declaring the migration complete. (Anthropic)
    • The Agent SDK billing split is operationally independent from the model cutover: credit pool defaults remain a single shared pool until per-user configuration is explicitly applied. Validate cost attribution routing before the billing cycle closes. (Anthropic)
  2. Summit runs June 15–18 — agent evaluation and Lakebase are the sessions to track
    • MLflow 3's trace-driven eval model is now the reference pattern for production regression testing of agentic apps; assess whether your current eval harness emits compatible spans or needs a migration of its own. (MLflow)
    • Lakebase GA on AWS changes the calculus for RAG pipelines that join structured metadata with vector search — a managed Postgres replacement that inherits lakehouse lineage is operationally meaningful for governed enterprise deployments. (Databricks)

Quick Links