What is ending
The opposite page to the preview radar. Runtimes losing support, models being retired, APIs going away, and behaviour changes with a date attached — on one countdown, nearest first.
4 inside ninety days
Inside thirty days
1This is not planning any more. If it is in production, it is this week’s work.
- in 2 days30 Oct 2026DeepSeek V4 Pro (0813) Model Serving Foundation Model APIs
DeepSeek V4 Pro (0813) is retired from Foundation Model APIs on 30 October 2026. Databricks names DeepSeek V4.1 Flash as the replacement.
Move workloads to DeepSeek V4.1 Flash, and follow the deprecated and retired models guidance for anything pinned to the 0813 version.
Already gone
4Here because older tutorials still recommend them, and because an incident on one of these starts with an upgrade.
- 166 days ago15 May 2026Llama 3.1 405B Model Serving
Retired from pay-per-token in February 2026 and from provisioned throughput in May.
The documented replacement is GPT OSS 120B.
- 104 days ago16 Jul 2026GPT-5.1 Codex Max, Codex Mini and GPT-5.2 Codex Model Serving
Three OpenAI coding models retired together in July 2026.
GPT-5.5 replaces the Max and 5.2 variants, GPT-5.4 Codex Mini replaces the Mini.
- 26 days ago2 Oct 2026Gemini 2.5 Flash and 2.5 Pro Model Serving
Both retire on the same day, Flash from pay-per-token and Pro from provisioned throughput.
Databricks points at Gemini 3.1 Pro or 3.5 Flash. Budget time for re-evaluating, not just renaming the endpoint.
- 19 days ago9 Oct 2026Claude Sonnet 4 on Foundation Model APIs Model Serving
Retired from pay-per-token serving. Requests to it will stop working.
Move to Claude Sonnet 4.6 and re-run your evaluations: a newer model is not a drop-in for prompt behaviour.
Dates come from the Databricks documentation and are re-checked by the agent. A runtime version is supported for roughly six months, and a long-term support version for three years from its general availability. Model retirements get at least three months of notice after the deprecation is announced.