Workers AI retires 24 models on 2026-11-20: the 410 that is not a 404, and the 5028 a retry policy drops

What changed

This is a catalog change, not a capacity or configuration change. Every ID below keeps working until 2026-11-20 and then fails permanently. No retry, quota change or gateway setting changes what a 410 means here, because the model is removed from the catalog rather than overloaded. Each item can be checked against the changelog and the model pages linked in the appendix.

• 24 models retire on 2026-11-20 (changelog dated 2026-11-05). The post says apps must remove references to the listed models “before November 20, 2026”, and that the date was moved from November 15. The list is 16 LLM IDs and 8 other models: text-to-speech, speech-to-text, image generation, vision and one embedding model. The full catalog is the only place that shows what remains.
• The failure shape is HTTP 410 with error code 5028. The changelog states that requests “will fail with an HTTP 410 Gone error and error code 5028”. The Workers AI errors table (last updated 2026-11-06) names the row Model deprecated, maps it to 410, and gives the message template ${model} was deprecated on ${date}. See the model catalog for alternatives. A 410 is a permanent answer: the same request will not start working later.
• Three similar-looking codes, three different bugs. 410 + 5028 means the model is retired. 404 + 3042 means the model name is invalid. 400 + 5007 means no such model or task, phrased as No such model ${model} or task. A typo, a wrong prefix and a retirement can all look like “the model call failed” if the code reads only the status line.
• The catalog already marks the retirement before it happens. Each deprecated model page carries a Planned deprecation badge and a Planned Deprecation | 11/20/2026 row while the model still answers. For example, @cf/meta/llama-3.2-1b-instruct and @cf/deepseek-ai/deepseek-v4-pro-0813 both show that row when reviewed on 2026-11-08. This is the cheapest way to confirm a single ID without sending a request.
• The LLM replacements are category recommendations, not a one-to-one map. The changelog recommends @cf/deepseek-ai/deepseek-v4.1-flash-1017 and @cf/zai-org/glm-5.3 for frontier needs, and @cf/google/gemma-4-26b-a4b-it and @cf/qwen/qwen3.8-27b for smaller needs. It does not say which old LLM maps to which new one, so the choice is a workload decision, not a rename. The eight non-LLM models do have explicit replacements, such as @cf/openai/whisper-tiny-en to @cf/deepgram/nova-3 and @cf/bytedance/stable-diffusion-xl-lightning to @cf/black-forest-labs/flux-2-klein-9b.
• The replacements differ in context, price, plan and defaults. Context windows run from 256,000 to 1,048,576 tokens, and two of the four recommended LLMs require the Workers Paid plan. The four also publish different reasoning-effort defaults, so an identical prompt can produce a different latency and token bill after a rename. The table below is copied from the model and pricing pages reviewed on 2026-11-08.
• Some workloads have no drop-in replacement at review time. Three deprecated LLMs are LoRA base models, and their replacement is a base model rather than the adapter you fine-tuned. The embedding model @cf/google/embeddinggemma-300m is mapped to @cf/google/embeddinggemma-2 with the note “available soon”. A team that fine-tuned or built an index on those models has a different problem from a team that only calls a chat model.

For scale: the deprecated @cf/meta/llama-3.2-1b-instruct lists a 60,000-token context at $0.027 / $0.201 per million tokens, and the deprecated @cf/deepseek-ai/deepseek-v4-pro-0813 lists 1,048,576 tokens at $1.32 / $3.96. Replacing a 1M-context model with a 256k-context model, or a cheap model with a paid-only one, is a product decision that a global search-and-replace hides. Cached-input prices are omitted here; read them on the pricing page before you size the bill.

Detect: are you affected?

Three read-only checks. Check 1 finds every copy of a retiring model ID in the repository. Check 2 proves what one call returns before and after the effective date. Check 3 confirms a single ID against the catalog metadata. Run them before changing a model name, a fallback list or a gateway config.

Check 1 — inventory every hardcoded model ID

The changelog is also published as Markdown, so the list can be extracted instead of retyped. The sed range picks only the deprecated-models section: the LLM block and the “Other models” block, where the first backticked ID on each line is the retiring model. Expect exactly 24 IDs.

# 1. Extract the retiring IDs from the published changelog.
curl -sS https://developers.cloudflare.com/changelog/post/2026-11-05-planned-model-deprecations/index.md \
  -o /tmp/retired.md
sed -n '/#### Models deprecated on November 20, 2026/,/For the full list of available models/p' /tmp/retired.md \
  | grep -E '^- `@cf/' \
  | sed -E 's/^- `([^`]+)`.*/\1/' \
  | sort -u > /tmp/retired-models.txt
wc -l /tmp/retired-models.txt

# 2. Find every file that still contains one of those exact IDs.
grep -RIlF -f /tmp/retired-models.txt . \
  --exclude-dir=node_modules --exclude-dir=.git --exclude-dir=.wrangler

# 3. Print the lines, so runtime copies can be separated from documentation copies.
grep -RInF -f /tmp/retired-models.txt . \
  --exclude-dir=node_modules --exclude-dir=.git --exclude-dir=.wrangler

Five places hold a model name in a typical AI site, and a repository scan only finds four of them: the front-end model picker, the backend route or environment variable, the evaluation and benchmark scripts, the fallback chain, and the AI Gateway model or alias field. The gateway one is configuration, not code: if the gateway was created in the dashboard, its model and fallback entries live only there, and a clean grep proves nothing about them.

Check 2 — probe one retiring ID against a control and a fake

The probe is deliberately three requests: one retiring ID from the list, one name that never existed, and one control model that is not on the list. The comparison is what distinguishes a retirement from a typo and from a credential problem. Set CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN first and keep the payload tiny; this command is read-only with respect to your application.

BASE="https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run"
probe () {
  printf '\n== %s\n' "$1"
  curl -sS -o /tmp/ai-probe.json -w 'HTTP %{http_code}\n' \
    -X POST "$BASE/$1" \
    -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \
    -H "Content-Type: application/json" \
    -d '{"messages":[{"role":"user","content":"reply with the single word ok"}]}'
  head -c 400 /tmp/ai-probe.json; echo
}
probe "@cf/meta/llama-3.2-1b-instruct"   # one ID from /tmp/retired-models.txt
probe "@cf/definitely-not-a-model"      # a name that never existed
probe "@cf/meta/llama-3.1-8b-instruct"  # a control that is not on the list

Check 3 — confirm one ID in the catalog metadata

Each model page is published as Markdown too. A deprecated model shows a planned deprecation row while it still answers, so this works before the effective date and costs one request per ID.

curl -sS https://developers.cloudflare.com/workers-ai/models/llama-3.2-1b-instruct/index.md \
  | sed -n '/Planned Deprecation/p'

Decision tree

inventory scan (check 1)
│
├─ no runtime copy ──► check dashboard + AI Gateway config
│                      ├─ clean ──► not affected: stop, do not buy
│                      └─ model present ──► affected: continue
│
└─ runtime copy ──► probe that ID (check 2)
                     │
                     ├─ 200 before 2026-11-20 ──► affected on schedule
                     │                            migrate before the date
                     ├─ 410 + 5028 ──► retirement live ──► FM-01, FM-02
                     ├─ 400 + 5007 / 404 + 3042 ──► wrong ID, not retirement
                     ├─ 403 + 5035 ──► replacement needs Paid ──► FM-07
                     └─ 401 / 403 (no 5035) ──► auth problem, out of scope

If checks 1 and 3 find nothing and the dashboard and gateway hold no retiring ID, stop: this runbook is not the right one, and buying it will not help. If the scan finds copies only in documentation, the same applies. The paid material starts once a runtime copy exists and the question is how to move it without changing behaviour silently.

Failure modes

Seven failures, each with the trigger, the documented error and the fix. Two are visible here.

Migration: what the paid chapter covers

The change is one date, but the work is per caller: each surface that stores a model name fails at a different moment, and some of them fail silently. The runbook splits the migration this way:

Appendix

• Changelog — Planned model deprecations on Workers AI (published 2026-11-05, reviewed 2026-11-08: 24 models, effective 2026-11-20, moved from 2026-11-15, HTTP 410 + code 5028, replacement list)
• Workers AI errors (last updated 2026-11-06, reviewed 2026-11-08: 5028 Model deprecated / 410; 3042 Invalid model ID / 404; 5007 No such model / 400; 5035 Workers Paid plan / 403)
• Workers AI pricing (last updated 2026-10-20, reviewed 2026-11-08: per-model unit prices and the list of models that require a paid billing method)
• Workers AI model catalog (reviewed 2026-11-08: the replacement IDs above are listed here; check the page for the exact current ID)
• deepseek-v4.1-flash-1017 model page (1,048,576 context, $0.30 / $1.20 per M, paid access required; reviewed 2026-11-08)
• glm-5.3 model page (1,048,576 context, $1.40 / $4.40 per M, paid access required; reviewed 2026-11-08)
• gemma-4-26b-a4b-it model page (256,000 context, $0.10 / $0.30 per M; reviewed 2026-11-08)
• qwen3.8-27b model page (262,144 context, $0.45 / $3.20 per M; reviewed 2026-11-08)
• llama-3.2-1b-instruct model page (deprecated: Planned Deprecation | 11/20/2026, 60,000 context, $0.027 / $0.201 per M; reviewed 2026-11-08)
• deepseek-v4-pro-0813 model page (deprecated: Planned Deprecation | 11/20/2026, 1,048,576 context, $1.32 / $3.96 per M; reviewed 2026-11-08)