Latest in AI
Workers AI November 20 Deprecations: Fix 410 and 5028 Before They Break Your App
Cloudflare's model cutoff needs more than new IDs: check paid access, fallback routes and embedding indexes before moving production traffic.
Cloudflare will stop serving a group of Workers AI models on November 20, 2026. Requests will return HTTP 410 Gone with error code 5028. If your application still references one of them, retries will not solve the underlying problem. You need a tested replacement route.
This is an approaching deadline, not a new announcement. Cloudflare's November 5 deprecation notice moved the original November 15 cutoff to November 20. The notice does not specify an exact cutoff hour, so I would not plan a last-minute switch around an assumed timezone.
Find the model calls that are easy to miss
Start with the complete affected-model list in the notice. It includes older DeepSeek, Kimi, Qwen, GLM, Llama and LoRA endpoints, alongside speech, image and embedding models. Matching the exact model ID matters more than recognizing a family name.
Search deployed configuration as well as source code. Check background jobs, environment-specific settings, evaluation scripts, cached routing decisions and fallback chains. A primary route can work while its backup still points to a retiring model, leaving the problem hidden until an unrelated failure activates it.
I would record each affected ID, the calling feature, its owner and the proposed replacement. Separate interactive requests from scheduled workloads so a successful chat test does not stand in for a transcription job that runs overnight.
Treat 5028 differently from a temporary capacity error
Cloudflare's error reference identifies 5028 as model deprecation. It separately lists invalid model IDs, access restrictions, timeouts and capacity errors. Those are different recovery paths.
When deprecation is the cause, repeatedly submitting the same model ID only repeats the failure. Route to an already validated alternative or return a clear unavailable result. Do not silently select an arbitrary model just because it accepts the request: its output may no longer meet the application's contract.
Log the selected model and error code without including private prompt content. That gives you enough evidence to distinguish a missed migration from a permission or capacity problem after deployment.
Check access before choosing an LLM replacement
Cloudflare recommends DeepSeek V4.1 Flash 1017 and GLM 5.3 for frontier workloads, with Gemma 4 26B A4B and Qwen3.8 27B for smaller ones. That is a shortlist, not a promise of identical behavior.
The DeepSeek replacement page explicitly requires Workers Paid or prepaid AI Gateway credits rather than standard Workers Free billing. Replacing a deprecated ID with a documented model can therefore exchange one error for an access problem if the account is not prepared.
Test the replacement through the same account and route used in production. Include the output shape your code expects, tool-call handling, streaming completion and a representative long input. Measure the request's actual usage before projecting spend. A successful short playground answer does not validate your application integration.
For image or speech replacements, review the new request and response schema independently. A suggested successor is not evidence that old parameters, formats or downstream parsing can remain untouched.
Resolve the EmbeddingGemma documentation mismatch
The deprecation notice names @cf/google/embeddinggemma-2 as the successor to @cf/google/embeddinggemma-300m, but still describes it as available soon. As reviewed on November 19, Cloudflare also has a full EmbeddingGemma 2 model page with usage examples and an API schema.
That page documents 768-dimensional output by default, with 128, 256 and 512 as shorter options. Its text example uses input and reads vectors from data[i].embedding. Review both sides of the adapter instead of changing only the model string. I have not made an authenticated inference call, so confirm access through your own account before relying on it for the cutoff.
The documentation also specifies task prefixes for retrieval queries and documents. Preserve those conventions consistently when generating a replacement corpus and its queries; otherwise, you are changing more than the checkpoint during evaluation.
Migrate the index and query path together
An embedding migration needs its own rollout. Cloudflare's Vectorize index guidance says dimensions and the distance metric are fixed at creation. If the replacement output size differs, create an appropriately configured index rather than sending incompatible vectors into the existing one.
Even matching dimensions do not establish that two models share a compatible embedding space. My proposed approach is to regenerate the document embeddings in a separate index, then test queries produced by the same replacement model. Preserve document IDs and metadata so retrieval results remain traceable to the original content.
Use a small set of questions with known useful results to check relevance, missing records and filtering. Only then switch the query encoder and index binding together. Keep the previous index for diagnosis, but remember that retaining it does not keep the deprecated query encoder available after shutdown.
The migration is complete when every production route has a validated destination, including fallbacks and retrieval. Removing one model name from the main request handler is only the beginning.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams turn API changes and migration risks into clear technical guidance developers can use.