Is Cloudflare AI Down?

If Prismix shows Cloudflare AI as "Operational" but you're having issues:

• Model not found 400 error — Workers AI model names are version-specific. Using a deprecated or misspelled model ID (e.g. @cf/meta/llama-2-7b vs the current @cf/meta/llama-3.1-8b-instruct) returns a 400. Check the Cloudflare AI model catalog for current identifiers before deploying.
• Rate limit 429 on the free tier — The Workers AI free tier has daily neuron limits per account. After exhausting the quota, every inference call returns HTTP 429 for the remainder of the day. Upgrade to Workers Paid or wait for the daily quota reset at midnight UTC.
• AI Gateway latency spike — Cloudflare AI Gateway is a reverse proxy that adds observability and caching. During high-traffic periods or when the Gateway region is under load, it can add hundreds of milliseconds of latency or time out before the inference even starts. Temporarily bypass AI Gateway by calling the Workers AI endpoint directly to confirm whether the Gateway or the model inference is the bottleneck.
• Wrangler binding not found locally — When developing with Wrangler locally (wrangler dev), the env.AI binding must be declared in wrangler.toml as [ai] binding = "AI". Missing this entry causes a runtime TypeError ("Cannot read properties of undefined") that looks like a platform outage but is actually a configuration error.
• GPU quota exceeded for the account — Enterprise and high-volume accounts can exhaust allocated GPU capacity during traffic spikes. Requests queue and eventually time out with a 503. Contact Cloudflare support to increase GPU quota, or add retry logic with exponential backoff in your Worker code.
• Response streaming broken in Workers preview — Workers AI streaming (SSE/ReadableStream) works differently in the local Wrangler preview versus the live Cloudflare edge. Some streaming-related bugs only appear in production. Use wrangler dev --remote to test against the actual Cloudflare network when debugging streaming issues.