Skip to content

Cloud · Analysis

Workers AI Retires 16 LLMs on Nov 20: Migration Guide

Cloudflare will switch off 16 Workers AI LLMs, including Kimi K2.6, GLM-5.2 and both DeepSeek V4 snapshots, on November 20, 2026. Here are the replacements, the cost impact and an 11-day migration plan.

Metarix Editorial · Nov 9, 2026 · 9 min read

Share

Cloudflare changelog card reading Planned model deprecations on Workers AI, dated 11.05
On this page

Cloudflare will switch off 16 large language models and 8 other models on Workers AI starting November 20, 2026. After that date, calls to any of them fail with HTTP 410 Gone and error code 5028, so a Worker that still names one breaks outright rather than degrading.

The list is not just legacy Llama 2 checkpoints. It includes Kimi K2.6, Kimi K2.7 Code, GLM-5.2, GLM-5.3 Flash and both DeepSeek V4 snapshots, the models many teams picked for agents this autumn. You have 11 days. Below is what is going, what to move to, what it does to your bill, and how to make the next catalog refresh a config change instead of an outage.

What Cloudflare announced

In a Workers AI changelog post dated November 5, Cloudflare said it is refreshing the model catalog "to make room for new releases and updated models" and asked developers to remove references to the listed models before November 20. The original cut-off was November 15; Cloudflare moved it back five days in the same post.

Two details matter for planning. First, the failure mode is hard: requests return 410 Gone with code 5028, not a silent redirect to a successor. Second, Cloudflare names replacements only for the non-LLM models (speech, image, vision and embeddings). For the 16 LLMs it simply recommends "newer models for the best token efficiency and throughput" and points to the pricing page, so choosing a successor is on you.

The 16 LLMs being retired, and where to move

The table lists every retiring LLM with its current price and a replacement we think fits the same job, drawn from models that are still in the Workers AI catalog and not on the deprecation list. The replacements are Metarix suggestions, not Cloudflare's. Prices as of November 9, 2026, from the Workers AI pricing page, per million tokens (input / output).

Scroll the table horizontally on smaller screens.

Retiring Workers AI LLMs and suggested replacements (prices per 1M tokens, input / output)
Retiring modelPrice nowSuggested replacementReplacement price
@cf/moonshotai/kimi-k2.6$0.95 / $4.00@cf/zai-org/glm-5.3$1.40 / $4.40
@cf/moonshotai/kimi-k2.7-code$0.95 / $4.00@cf/zai-org/glm-5.3$1.40 / $4.40
@cf/zai-org/glm-5.2$1.40 / $4.40@cf/zai-org/glm-5.3$1.40 / $4.40
@cf/zai-org/glm-5.3-flash$0.15 / $0.50@cf/zai-org/glm-4.7-flash or @cf/deepseek-ai/deepseek-v4.1-flash-1017$0.06 / $0.40 or $0.30 / $1.20
@cf/deepseek-ai/deepseek-v4-pro-0813$1.32 / $3.96@cf/zai-org/glm-5.3 (quality) or deepseek-v4.1-flash-1017 (cost)$1.40 / $4.40 or $0.30 / $1.20
@cf/deepseek-ai/deepseek-v4-flash-0731$0.44 / $1.32@cf/deepseek-ai/deepseek-v4.1-flash-1017$0.30 / $1.20
@cf/nvidia/nemotron-3-120b-a12b$0.50 / $1.50@cf/openai/gpt-oss-120b$0.35 / $0.75
@cf/qwen/qwq-32b$0.66 / $1.00@cf/openai/gpt-oss-120b or @cf/qwen/qwen3.8-27b$0.35 / $0.75 or $0.45 / $3.20
@cf/deepseek-ai/deepseek-r1-distill-qwen-32b$0.497 / $4.881@cf/openai/gpt-oss-120b$0.35 / $0.75
@cf/qwen/qwen3-30b-a3b-fp8$0.051 / $0.335@cf/google/gemma-4-26b-a4b-it or @cf/zai-org/glm-4.7-flash$0.10 / $0.30 or $0.06 / $0.40
@cf/meta/llama-3.1-8b-instruct-fp8$0.152 / $0.287@cf/meta/llama-3.1-8b-instruct-fp8-fast$0.045 / $0.384
@cf/meta/llama-3.2-1b-instruct$0.027 / $0.201@cf/ibm-granite/granite-4.0-h-micro$0.017 / $0.112
@cf/meta/llama-3.2-3b-instruct$0.051 / $0.335@cf/ibm-granite/granite-4.0-h-micro$0.017 / $0.112
@cf/meta/llama-2-7b-chat-hf-loraLoRA base@cf/mistral/mistral-7b-instruct-v0.2-lora (retrain adapter)See pricing page
@cf/google/gemma-2b-it-loraLoRA base@cf/mistral/mistral-7b-instruct-v0.2-lora (retrain adapter)See pricing page
@cf/google/gemma-7b-it-loraLoRA base@cf/mistral/mistral-7b-instruct-v0.2-lora (retrain adapter)See pricing page
Retiring list: Cloudflare Workers AI changelog, Nov 5, 2026. Prices: Workers AI pricing page, checked Nov 9, 2026. Replacements are Metarix suggestions based on model class and the live catalog; test on your own prompts before switching.

A few of these pairings are judgment calls, so here is the reasoning. GLM-5.3 is the closest remaining match for Kimi and GLM-5.2 workloads: Cloudflare's catalog describes it as Z.ai's flagship agentic coding model with a 1M-token context window, function calling and structured outputs. DeepSeek V4.1 Flash 1017 is a straight upgrade path for the V4 Flash snapshot and is cheaper on both input and output. For small reasoning models, gpt-oss-120b now undercuts QwQ-32B and the R1 distill by a wide margin on output tokens.

The non-LLM models have official replacements

For speech, image, vision and embeddings, Cloudflare did name successors. These are the easiest swaps, though output formats and voices differ, so check downstream parsing.

Scroll the table horizontally on smaller screens.

Retiring non-LLM Workers AI models and Cloudflare's named replacements
Retiring modelCloudflare's replacementTask
@cf/myshell-ai/melotts@cf/deepgram/aura-2-enText-to-speech
@cf/stabilityai/stable-diffusion-xl-base-1.0@cf/black-forest-labs/flux-2-klein-9bImage generation
@cf/runwayml/stable-diffusion-v1-5-inpainting@cf/black-forest-labs/flux-2-klein-9bImage editing
@cf/bytedance/stable-diffusion-xl-lightning@cf/black-forest-labs/flux-2-klein-9bFast image generation
@cf/lykon/dreamshaper-8-lcm@cf/black-forest-labs/flux-2-klein-9bImage generation
@cf/openai/whisper-tiny-en@cf/deepgram/nova-3Speech-to-text
@cf/llava-hf/llava-1.5-7b-hf@cf/moondream/moondream3.1-9B-A2BImage-to-text
@cf/google/embeddinggemma-300m@cf/google/embeddinggemma-2 (Cloudflare says "available soon")Embeddings
Source: Cloudflare Workers AI changelog, Planned model deprecations, Nov 5, 2026.

The embeddings row deserves a warning. The named successor was not yet available when Cloudflare published the notice, and switching embedding models means re-embedding your corpus, because vectors from different models are not comparable. If you store EmbeddingGemma 300M vectors in Vectorize or another vector database, plan the re-index now rather than on November 19.

Why this list is more disruptive than it looks

Catalog refreshes usually prune models nobody has touched in a year. This one removes several that were current. GLM-5.3 Flash is notable: Cloudflare's own catalog still describes it as outperforming GLM-5.2 "at one-tenth the price", yet it is on the retirement list while full GLM-5.3 stays. Both Kimi models go, and the models catalog we checked lists no other Kimi model; the pricing page still shows @cf/moonshotai/kimi-k2.5, so confirm availability in your account before treating it as an escape hatch.

Cloudflare gave no reason for individual cuts. The practical reading for builders is that on any serverless host, a model ID is a lease, not a purchase. We made the same point after Together repriced Qwen3.7-Max: third-party catalogs move on their own schedule, and your architecture should assume it.

What the switch does to your bill

Here is an illustrative month of 100 million input tokens and 10 million output tokens, with no caching, for the most common migrations. These are list-price calculations by Metarix, not measurements.

Scroll the table horizontally on smaller screens.

Illustrative monthly cost: 100M input + 10M output tokens, no caching
MigrationBeforeAfterChange
Kimi K2.6 to GLM-5.3$135.00$184.00+36%
GLM-5.2 to GLM-5.3$184.00$184.000%
DeepSeek V4 Flash 0731 to V4.1 Flash 1017$57.20$42.00-27%
DeepSeek V4 Pro 0813 to V4.1 Flash 1017$171.60$42.00-76%
Nemotron 3 Super to gpt-oss-120b$65.00$42.50-35%
GLM-5.3 Flash to GLM-4.7 Flash$20.00$10.00-50%
Qwen3 30B A3B to Gemma 4 26B A4B$8.45$13.00+54%
Calculated by Metarix from Workers AI list prices, checked Nov 9, 2026. Cached-input discounts, which GLM and DeepSeek models offer, would change these numbers.

The pattern: Kimi users pay more to stay on a comparable agentic model, while DeepSeek and Nemotron users can save money if a cheaper model holds up on their evals. Moving from V4 Pro to V4.1 Flash is the biggest saving on paper, but it is a step down in model class, so measure task success, not just tokens. Our guide to cost per successful task explains why a cheaper model that retries more can cost more overall.

A migration checklist for the next 11 days

  • Find every reference. Search code, config, environment variables and prompt templates for the retiring IDs. Remember model IDs also hide in wrangler config vars, database rows and feature flags.
  • Check live traffic, not just code. If you route Workers AI through AI Gateway, filter its logs by model to catch calls from services you forgot about.
  • Run your evals on the replacement. Use a few hundred real prompts per workload. Agentic flows need tool-call accuracy and JSON validity checks, not just answer quality.
  • Watch paid-only models. Cloudflare's pricing page says GLM-5.3, DeepSeek V4.1 Flash 1017 and the other frontier models require Workers Paid or prepaid AI Gateway credits. A free-plan project moving off Kimi or GLM-5.2 will not get around that.
  • Retrain LoRA adapters. Adapters trained on Llama 2 7B or Gemma 2B/7B bases will not load on a different base model.
  • Deploy before November 19. Leave a day to roll back if a replacement misbehaves under production load.

A quick way to find hard-coded IDs in a repository (illustrative; adjust the file types for your project):

# Illustrative: list files that reference a retiring Workers AI model
cat > retiring.txt <<'EOF'
kimi-k2.6
kimi-k2.7-code
glm-5.2
glm-5.3-flash
deepseek-v4-pro-0813
deepseek-v4-flash-0731
nemotron-3-120b-a12b
qwq-32b
deepseek-r1-distill-qwen-32b
qwen3-30b-a3b-fp8
llama-3.1-8b-instruct-fp8"
llama-3.2-1b-instruct
llama-3.2-3b-instruct
llama-2-7b-chat-hf-lora
gemma-2b-it-lora
gemma-7b-it-lora
EOF
grep -rnF -f retiring.txt --include='*.ts' --include='*.js' --include='*.json' --include='*.toml' --include='*.py' .

The Llama 3.1 line ends with a closing quote on purpose, so it does not flag the llama-3.1-8b-instruct-fp8-fast variant, which is not being retired. We ran this pattern against a small sample file in this session to check it matches the retiring IDs and skips the fast variant; it will not catch IDs built from string concatenation.

Make the next refresh a config change

The best long-term fix shipped the day before the deprecation notice. On November 4, Cloudflare added model aliases to AI Gateway: a named pool of models, such as custom/support-fast, that your application calls by one name. The gateway tries the models in the order you set and falls back to the next one if a request fails, or uses smart routing to pick a model per request. Because your code only references the alias, you can swap models without redeploying, according to the Cloudflare changelog.

Per Cloudflare's example, you call an alias through the REST API by using its name as the model and sending the cf-aig-gateway-id header to pick the gateway where the alias lives:

POST /accounts/{account_id}/ai/run
cf-aig-gateway-id: <your-gateway-id>

{
"model": "custom/support-fast",
"input": {
"messages": [{ "role": "user", "content": "Reply with a short greeting." }]
}
}

If you are not ready for aliases, at least move the model ID into configuration and handle the retirement error explicitly. The sketch below is illustrative and was not run in this session; check how your client surfaces the 410 status before relying on it.

// Illustrative Worker: model IDs from config, with one fallback
export default {
async fetch(req, env) {
const messages = [{ role: "user", content: "Summarize this ticket" }];
try {
return Response.json(await env.AI.run(env.PRIMARY_MODEL, { messages }));
} catch (err) {
// Retired models answer 410 Gone with error code 5028
console.error("primary model failed", env.PRIMARY_MODEL, String(err));
return Response.json(await env.AI.run(env.FALLBACK_MODEL, { messages }));
}
},
};

Aliases also fit a broader gateway strategy. If you want fallbacks that span providers, not just Workers AI models, our AI gateways comparison covers how Cloudflare's gateway stacks up against Vercel's and OpenRouter.

Verdict

Treat November 20 as a hard deadline. If you run Kimi K2.6, Kimi K2.7 Code or GLM-5.2 for agents on Workers AI, GLM-5.3 is the lowest-friction move, at the same price as GLM-5.2 and about a third more than Kimi on our example workload. If you run either DeepSeek V4 snapshot, test DeepSeek V4.1 Flash first, because it is cheaper and may be good enough. Small-model users should look at gpt-oss-120b, Gemma 4 and Granite 4.0 Micro.

Then fix the root cause. Put an AI Gateway alias or, at minimum, a config-driven model ID with a fallback in front of every model call. Cloudflare has now shown that even recent models can leave the catalog with about two weeks' notice, and the next refresh will be cheaper if swapping a model never needs a deploy.

Sources

Tools covered in this article

Frequently asked questions

When do the deprecated Workers AI models stop working?
Cloudflare says the listed models stop returning responses starting November 20, 2026. The date was originally November 15. After retirement, requests fail with HTTP 410 Gone and error code 5028.
Which Workers AI LLMs are being deprecated?
Sixteen LLMs, including Kimi K2.6, Kimi K2.7 Code, GLM-5.2, GLM-5.3 Flash, DeepSeek V4 Pro 0813, DeepSeek V4 Flash 0731, Nemotron 3 Super, QwQ-32B, the DeepSeek R1 distill of Qwen 32B, Qwen3 30B A3B FP8, Llama 3.1 8B Instruct FP8, Llama 3.2 1B and 3B, and three LoRA base models.
What should I replace Kimi K2.6 with on Workers AI?
GLM-5.3 is the closest remaining agentic coding model in the Workers AI catalog, at $1.40 input and $4.40 output per million tokens as of November 9, 2026, versus $0.95 and $4.00 for Kimi K2.6. Cloudflare did not name an official replacement, so test on your own prompts.
Will Cloudflare automatically redirect retired models to new ones?
No. Cloudflare's notice says retired models return a 410 Gone error rather than being redirected. To swap models without redeploying, you can use AI Gateway model aliases, which Cloudflare added on November 4, 2026.

Share

Keep Reading