Cloudflare will retire 24 Workers AI models on November 20

The extended deadline is a hard stop. Image request formats and embedding indexes may need more work than a model name change.
Cloudflare plans to retire 24 Workers AI model endpoints on November 20, 2026, cutting off responses from applications that still call them. The company’s November 5 notice covers 16 language models and eight models for other tasks. It extends an earlier November 15 deadline by five days. Cloudflare’s retirement notice
Failed calls will return HTTP 410 Gone with Cloudflare error code 5028. The company’s error reference identifies that combination specifically as a retired model, separate from a timeout or temporary capacity problem. Retrying the same endpoint will not fix a retirement. Workers AI error reference
Which models are affected
The list includes DeepSeek R1 Distill Qwen 32B, QwQ 32B, Kimi K2.6 and K2.7 Code, GLM 5.2 and GLM 5.3 Flash, several Llama models, and three LoRA-oriented endpoints. Cloudflare recommends DeepSeek V4.1 Flash 1017 or GLM 5.3 for frontier workloads, and Gemma 4 26B A4B or Qwen 3.8 27B for smaller-model needs. These are category-level recommendations, not one-to-one compatibility guarantees. Full model IDs and recommendations
That distinction matters for applications built around structured output, tool calls, streaming, or adapters. A replacement should be tested against the behavior the application actually depends on, including malformed inputs and failure handling, before traffic moves.
Image migrations need a request format check
Cloudflare directs four retiring image endpoints, including Stable Diffusion XL and Stable Diffusion 1.5 Inpainting, to FLUX.2 [klein] 9B. Retirement notice
FLUX.2 [klein] 9B uses multipart form data even for a text-only prompt. Its inference process is fixed at four steps, and the steps parameter cannot be adjusted. Cloudflare’s examples also show how to preserve the multipart boundary when calling through a Workers binding. Those details make a search-and-replace of the model string an incomplete migration for clients built around a different request shape. Cloudflare’s FLUX.2 integration guide
Teams using masks or other inpainting-specific controls should separately test their editing workflow. A suggested replacement model does not establish that every input field or editing behavior carries over.
Embedding migrations can reach into stored data
The notice also retires @cf/google/embeddinggemma-300m and names @cf/google/embeddinggemma-2 as its successor, marked “available soon” in the announcement. Retirement notice
An embedding model converts content into the vectors used for semantic retrieval. Switching only the query encoder can leave a system comparing new query vectors with documents encoded by the old model. Pinecone’s explanation of its own integrated embedding system illustrates the underlying requirement: it uses the same model for documents and queries, and re-embeds imported text into the destination index. Pinecone’s embedding architecture
For teams changing embedding models, a prudent migration is to build a separate index, re-embed the source documents, and compare retrieval results on representative queries before switching readers. Matching vector dimensions alone should not be treated as proof that two models produce compatible representations. Confirm the successor’s availability and supported configuration before committing the cutover.
Check deployed traffic and the fallback path
A useful first pass is to compare Cloudflare’s exact retiring IDs against deployed configuration, application code, scheduled jobs, and fallback routing. Then use request logs or equivalent telemetry to establish which matches receive traffic. A name in a model catalog identifies something to investigate; it does not establish live use or an impending outage.
Test a retired-model error deliberately in staging. The application should stop retrying that endpoint and either use a tested alternative or return a clear failure. Check fallback models too: an infrequently used route can remain invisible until the primary model fails.
Finally, review the replacement’s billing requirements. Cloudflare’s pricing page says both recommended frontier options, GLM 5.3 and DeepSeek V4.1 Flash 1017, require a paid billing method through Workers Paid or prepaid AI Gateway credits. Workers AI pricing
The deadline leaves teams with a concrete release criterion: affected production paths should use tested, available replacements before November 20, with retrieval quality, request handling, and fallback behavior verified rather than assumed.
Sources and citations
Revision history
Published .