Releasesby GlyphDeepSeek V4 Flash Vision Experimental now available on AI Gateway
DeepSeek V4 Flash with vision is now available on AI Gateway. This model is an experimental version that accepts images alongside text. You can ask it to describe a picture, read text out of a screenshot, or work through a chart in the same request as your prompt. DeepSeek V4 Flash Vision Experimental now available on AI Gateway.
Also shipped Aug 21, 2026Vercel in August 2026 →
Version targeting for Cortex Agent and Cortex Analyst evaluations (General availability)
Cortex Agent and Cortex Analyst evaluations now let you control the versioned configuration used for scoring: Version targeting pins the configuration under evaluation, not the agent behavior that gets evaluated: each run invokes the agent again and scores a newly generated trace.
CoCo automations in CLI and Snowsight (Preview)
CoCo automations in the CoCo CLI and CoCo in Snowsight are now available in public preview. Automations let you schedule recurring, unattended CoCo runs that execute in a Snowflake-managed sandbox. Each run creates a Cortex thread that you can inspect and continue in CoCo. Automations created from one surface can be monitored and managed from the other.
Upcoming end of life for the Supervisor API
The Supervisor API (Beta) reaches end of life on September 30, 2026. After that date, the API is no longer available. Databricks recommends migrating to custom agents on Databricks Apps . See Supervisor API (Beta) (deprecated) .
Consumer access to query Unity Gateway services will become generally available
Querying Unity Gateway model services and model provider services with Consumer access is currently opt-in through the Consumer access to Unity Gateway Public Preview . When this capability becomes generally available, consumer access users will be able to query these services without the preview enabled.
Monitor all AI activity with the unified trace table (Beta)
The unified trace table is in Beta . It captures every request and response across your Unity Gateway services into one Unity Catalog table in OpenTelemetry format, giving you a single place to monitor, debug, secure, and audit all AI activity. See Monitor all AI activity using the unified trace table .
Default Python package repositories are GA for Lakeflow pipelines and classic compute
Configuring default Python package repositories is now generally available for Lakeflow pipelines and for classic compute created via the UI and API. Workspace admins can set private or authenticated registries as the workspace default for these workloads. See Configure default package repositories .
More from AI SDK and AI GatewayEvery AI SDK and AI Gateway release →
Vercel Agent is now available in Slack code channels
Vercel Agent now works in Slack code channels, a new kind of channel launched today for working with a coding agent. Anyone in the channel can follow the work, give Agent new instructions, and review the code it writes. Choose Create a code channel from the Slack sidebar, select Vercel as the agent, and describe what you need to get started.
Fish Audio models now available on Vercel AI Gateway for free
Fish Audio 's audio models are now available on AI Gateway. To celebrate the launch, every Fish Audio model is free on AI Gateway for the next 30 days, through September 18.
Exa joins the Vercel AI Gateway and Agent Marketplace
Exa is now available on the Vercel AI Gateway and Agent Marketplace as a native integration. Exa's neural search engine grounds models in current information. Exa search is now a built-in tool on AI Gateway , which runs the search for you without an Exa account of your own.
Vercel Sandbox is now globally available
Vercel Sandbox now runs globally, starting with four regions: iad1 (Washington, D.C.), sfo1 (San Francisco), cle1 (Cleveland), and cdg1 (Paris). iad1 remains the default. Support for all Vercel regions is coming soon. Choose a region close to the databases, object storage, and other services your sandboxes access to reduce latency.
The Vercel Sandbox CLI is now more agent-friendly
The Sandbox CLI version 4.0.1 ships four output and error improvements: New sandboxes point users towards how to connect into them. A sandbox not found error shows where to look for existing sandboxes. Common mistakes, like typing a command after sh , redirect to the command that was intended.
Cline is now available in the AI SDK harness layer
The AI SDK harness layer lets you run established coding-agent runtimes through one unified interface, so you can switch runtimes without changing your application code. Today we are adding Cline, which runs through the same HarnessAgent interface as other supported harness.
GLM 5.3 now available on AI Gateway
GLM 5.3 from Z.ai is now available on AI Gateway. GLM 5.3 has improvements vs. GLM 5.2 at complex software engineering and at agent tasks that run across many steps, and it reaches those results while producing fewer output tokens than GLM 5.2 did at the same effort level.
MiniMax M3 and M2.7 are free on AI Gateway
MiniMax M3 and M2.7 are free on AI Gateway via GMI Cloud through Sunday, September 6. Use minimax/minimax-m3-free or minimax/minimax-m2.7-free to route requests to GMI Cloud. These model IDs will return an error after the free period ends.