Is AI21 Labs Down?

If Prismix shows AI21 Labs as "Operational" but you're having issues:

• Jamba 256k context window truncating at 128k — the context_length parameter may not be respected if you're calling an older Jamba model variant. Verify you are using jamba-1.6-large or newer, and that your API plan supports the full 256k context window. Older model slugs cap at 128k regardless of what you pass.
• Jurassic-2 API endpoint deprecated — AI21 Labs has sunset the Jurassic-2 (j2-*) endpoints in favor of the Jamba API. If you're receiving 404 or 410 errors on /v1/complete, migrate to the chat completions endpoint under the Jamba model family.
• API key invalid — must use studio key not playground key — AI21 Labs has two separate key namespaces: keys generated in the AI21 Studio console (studio.ai21.com) and keys from the older playground. Only Studio API keys work with the production API. Regenerate your key from the Studio dashboard and update your environment variables.
• chat_only mode returning non-chat format — if you set chat_only: true but the response is still in completion format, you may be hitting an older cached model version. Explicitly pass model: "jamba-1.6-large" and use the /v1/chat/completions endpoint rather than the legacy completion path.
• Rate limit 429 on free tier (5 req/min) — the AI21 Labs free tier enforces a hard limit of 5 requests per minute. Burst traffic from parallel application calls will reliably hit this limit. Implement request queuing with a 13-second minimum interval between calls, or upgrade to a paid plan for higher rate limits.
• Streaming response not finishing (connection reset before final token) — streaming via the Jamba API can be interrupted if the client closes the connection before receiving the stop token, or if a network proxy buffers the entire stream before forwarding. Ensure your HTTP client has streaming enabled (e.g., stream=True in Python) and set a generous read timeout of at least 120 seconds for long outputs.