Cloudflare sets November 15 deadline for Workers AI model changes
Cloudflare has told Workers AI developers to remove references to 24 models before November 15, 2026. Its November 5 notice covers 16 language models and eight models for other tasks. Teams should check their deployed model identifiers now, then test replacement behavior, access and costs before changing production traffic.
Find the affected calls first
The list includes older DeepSeek, Kimi, GLM, Qwen, Llama and Gemma entries, alongside speech, image and vision models. The specific deployment name matters. Check the complete deprecation list rather than assuming every model in an affected family is being retired.
Start with application configuration, saved workflows, fallback routes and background jobs. A model that is rarely called may still be the one a service depends on when its usual provider fails. Record which feature uses each identifier and who will verify its replacement.
The announcement gives a calendar date without an exact cutoff time or timezone. Completing the change ahead of that date leaves room to catch problems while the current setup is still available for comparison. ByteForward has not run these migrations or independently tested the replacements.
Compare access and reasoning settings
Cloudflare points to DeepSeek V4.1 Flash 1017 and GLM 5.3 for frontier needs, and Gemma 4 26B A4B and Qwen 3.8 27B for smaller workloads.
The DeepSeek listing requires Workers Paid billing or prepaid AI Gateway credits. It documents a 1,048,576 token context window, vision and function calling. Its reasoning settings include none, low, high and max, with high as the default.
GLM 5.3 has the same paid access requirement and listed context length. Its reasoning cannot be disabled. The documentation maps compatibility values such as none and medium to max. A shared application setting could therefore produce very different behavior after a model change.
The smaller choices differ too. Gemma 4 26B A4B lists 256,000 context tokens. Qwen 3.8 27B lists 262,144 and defaults to xhigh reasoning. Both list vision and function calling, but those labels do not establish equivalent results on a particular task.
Recalculate the bill with realistic outputs
Cloudflare’s current pricing table gives these USD rates per million tokens:
| Model | Input | Output |
|---|---|---|
| DeepSeek V4.1 Flash 1017 | $0.30 | $1.20 |
| GLM 5.3 | $1.40 | $4.40 |
| Gemma 4 26B A4B | $0.10 | $0.30 |
| Qwen 3.8 27B | $0.45 | $3.20 |
These are standard input rates. Cached input rates, where supported, are separate. The table is a starting point for a cost estimate, not a forecast of cost per successful task.
Run the same representative requests through the candidates and record total tokens, retries, latency and whether the result was usable. A model with a lower token price can still require more output or more attempts. Conversely, a more expensive response may be worthwhile if it reliably completes work that previously failed. Those comparisons need application evidence.
Speech and image replacements need separate checks
For speech recognition, Nova 3 lists different transport prices: $0.0052 per audio minute over regular HTTP and $0.0092 over WebSocket. Its schema also includes controls for speaker changes, language, punctuation and interim results. Check which settings the application actually depends on and how its response parser handles the returned data.
For a transcription shortlist, test the accents, microphones, background noise and specialist vocabulary your users bring. ByteForward’s guide to the expanded Open ASR Leaderboard explains why a headline average can hide relevant differences.
Text to speech has another billing unit. Aura 2 English costs $0.03 per 1,000 characters and exposes voice, encoding and sample rate controls. Estimate spoken output from real text rather than comparing its character rate directly with an audio minute rate.
The FLUX.2 klein 9B listing documents a multipart input object and a Base64 image result. An image pipeline should test upload handling, decoding and the required editing behavior before routing user requests to it.
Leave time for the embedding change
Cloudflare names EmbeddingGemma 2 as the replacement for EmbeddingGemma 300M, but the notice still describes it as coming soon. Do not assume that migration route is already usable.
For an embedding migration, plan to validate retrieval quality and index compatibility separately. Keep a known working route available while testing the new setup, and confirm current availability before committing to the switch. The immediate goal is continuity: every affected application should have a tested path that survives the deadline.
