Back Home

雲端 AI/模型生命週期

OCI Retires Ten Legacy Grok Endpoints on the Same Day; Existing Deployments Must Migrate to `xai.grok-4.3`

Oracle’s August 15 retirement date for Grok 3, Grok 4, and multiple Fast, Mini, and Code endpoints has passed. This is not a transparent model-alias switch: OCI users must change the model ID and revalidate inference parameters, image limits, and regional availability.

Aeroprints.com · CC BY-SA 3.0 · Image source
zh-Hant

Oracle’s OCI Generative AI retirement table shows that ten xAI on-demand models reached their end-of-service date on August 15, including Grok 3, Grok 3 Mini, the Fast variants of both, Grok 4, Grok 4 Fast, Grok 4.1 Fast, and Grok Code Fast 1; the documentation assigns the same date to certain duplicate version listings. Oracle specifies Grok 4.3 as the common replacement, with the OCI model ID `xai.grok-4.3`. The legacy model pages explicitly warn that the models will become inaccessible after retirement, creating an immediate outage risk for inference, RAG, or agent workflows with hard-coded legacy IDs.

Migration involves more than replacing a single name. Grok 4.3 on OCI provides a one-million-token context window, supports only on-demand mode, and is accessible through the Console, Inference API, and CLI. The image API accepts only base64-encoded JPEG or PNG input, and each image must represent between 256 and 1,792 tokens after conversion; text and all images still share the overall context limit. Oracle also cautions that not every model is available in every listed region. Because Grok on OCI is managed by xAI, teams should recheck data-path and regional requirements.

Application teams should use a fixed regression suite to compare the old and new models’ tool-calling schemas, structured outputs, refusal behavior, latency, and token usage. In particular, they should not assume that Grok 4.3 preserves the cost and latency characteristics of the legacy Fast or Mini variants. xAI states that the 4.3 interface supports `none`, `low`, `medium`, and `high` reasoning effort, but the parameters accepted by the cloud wrapper should still be verified against OCI documentation and through testing. Oracle’s model catalog may temporarily retain entries marked “Deprecated”; this does not mean that retired endpoints retain any formal availability guarantee.

Sources

  1. Model Retirement Dates (On-Demand Mode)
  2. Grok on Amazon Bedrock