Workers AI shuts down 24 models on November 20, and calls start returning error 410

On November 5, 2026, Cloudflare published in its developer changelog the list of models that stop working on Workers AI on November 20. The original date was November 15 and was pushed back five days. From the cutoff, any request to one of these models gets HTTP 410 Gone with error code 5028, whose message says the model was deprecated and points to the catalog. The changelog mentions no transition period after that date.

Which models go and what replaces them

There are 16 text models and 8 in other modalities. For the text models, Cloudflare does not name a one-to-one replacement. It recommends two groups: @cf/deepseek-ai/deepseek-v4.1-flash-1017 and @cf/zai-org/glm-5.3 as larger models, and @cf/google/gemma-4-26b-a4b-it and @cf/qwen/qwen3.8-27b as smaller options.

Deprecated text models:

• @cf/deepseek-ai/deepseek-r1-distill-qwen-32b, @cf/deepseek-ai/deepseek-v4-pro-0813 and @cf/deepseek-ai/deepseek-v4-flash-0731
• @cf/moonshotai/kimi-k2.6 and @cf/moonshotai/kimi-k2.7-code
• @cf/zai-org/glm-5.2 and @cf/zai-org/glm-5.3-flash
• @cf/qwen/qwq-32b and @cf/qwen/qwen3-30b-a3b-fp8
• @cf/nvidia/nemotron-3-120b-a12b
• @cf/meta/llama-3.1-8b-instruct-fp8, @cf/meta/llama-3.2-1b-instruct and @cf/meta/llama-3.2-3b-instruct
• the LoRA base models @cf/meta-llama/llama-2-7b-chat-hf-lora, @cf/google/gemma-2b-it-lora and @cf/google/gemma-7b-it-lora

The rest have a named replacement:

Who needs to act

Any Worker, function or backend that calls one of these identifiers, through the env.AI binding or the REST API. The identifier usually lives in a constant or an environment variable, which is where the search starts. Dormant projects, prototypes that became production features and integrations built by an agency that has already handed over the work are the easiest to forget.

Swapping one text model for another is rarely just changing a string. Prompts tuned for one model behave differently on another, and the length and shape of the answer can change what the application does next. The replacement image, speech and transcription models come from other vendors, with their own parameters.

Embeddings are the most delicate case. Vectors produced by different models are not comparable, so replacing embeddinggemma-300m means regenerating the vectors for the whole indexed collection, not only for new queries. And the suggested replacement was not yet available on the day of the notice.

What to do before November 20

1. Search code and configuration for the identifiers above, including repositories that have not had a commit in months.
2. For each use, choose the replacement and run a small set of real inputs through both models before switching. Compare output, latency and per-token cost on the pricing page.
3. Handle error 5028 explicitly, with a clear message and an alert to the team. A 410 treated as a generic failure can go unnoticed for days.
4. If you use embeddinggemma-300m: plan the reindexing and decide what to do if embeddinggemma-2 does not arrive before the cutoff.

A notice with a fifteen-day deadline is part of the price of using a managed model catalog: the vendor sets the date of the switch. I wrote about that cost in A tool's cost includes the work to replace it.

What is still unclear

The changelog does not say whether there will be another extension, or whether batch requests or AI Gateway routes pointing to these models get any different treatment. It also does not say what happens to LoRA adapters trained on the three deprecated base models, or when embeddinggemma-2 becomes available. Anyone who depends on one of these points should check Cloudflare's model catalog and changelog before the cutoff.

If your company is weighing something similar, reach me on LinkedIn or WhatsApp.

References

• Cloudflare Changelog: Planned model deprecations on Workers AI (November 5, 2026)
• Cloudflare Docs: Workers AI error codes
• Cloudflare Docs: Workers AI model catalog
• Cloudflare Docs: Workers AI pricing