Hi team, We’re currently experiencing this error during some live calls: "An error occurred in this call: call.in-progress.error-vapifault-azure-openai-429-exceeded-quota" We are using Vapi’s built-in OpenAI provider with model GPT-5.2. Could you please confirm: Are we hitting any workspace rate limits or concurrency limits? Is this related to our current plan’s quota/credit cap? Are there specific RPM/TPM limits applied per assistant or workspace? The calls are ending immediately when this happens, so we’d like to understand whether this is traffic-related or plan-related. Thanks in advance.

Thanks for reaching out happy to help with this! The error you're seeing is a 429 Quota Exceeded from Azure OpenAI, meaning the requests are hitting the rate or quota limits on Vapi's shared Azure OpenAI provider. This is typically traffic-related (too many concurrent calls exhausting shared RPM/TPM limits), though plan-level credit caps can also trigger it. Could you let us know roughly how many concurrent calls you're running when this happens? That'll help us pinpoint whether it's a concurrency issue or a plan quota cap.

Thanks for the response. To clarify from our side: We are splitting campaigns specifically to keep concurrency low. On Mar 2, 2026 – 8:30 AM CST, the campaign had only 6 total calls, and 1 call received the 429 exceeded quota error. Later on Mar 2, 2026 – 9:26 AM CST, another campaign had 6 total calls, and 2 calls received the same error. So concurrency was limited to a maximum of 6 calls within the campaign, yet the 429 error still occurred. During these same periods, some customers also experienced brief dead air/silence from the assistant before the call dropped. Based on this, could you please clarify whether this is due to shared Azure OpenAI RPM/TPM limits across workspaces, internal throttling on the shared provider, or a plan-level quota restriction on our account?

---- Hi, This is a Vapi-side issue. It's not related to your plan or workspace limits. Two ways to avoid it in the meantime: 1. Switch to GPT-4.1 - higher throughput, no rate limit issues 2. Bring your own OpenAI API key - isolates you from shared quota entirely We're working on a fix on our end. Sorry for the disruption! Best, Vapi Support

With calls triggering a 429, this almost certainly points to shared Azure OpenAI RPM/TPM pool throttling across workspaces not your concurrency or plan quota. Vapi's shared provider pools capacity across all users, so even low-volume campaigns can get squeezed during peak times. I've dealt with this exact issue before and can help you get it sorted properly. If you're open to it, I'd love to jump on a quick call to walk you through the fix and potentially help on an ongoing basis if you need someone with Vapi/AI call infrastructure experience. @User

Hi, thanks for the update. Since we plan to switch to bringing our own OpenAI API key, we’d like to understand the expected throughput requirements from Vapi’s side so we can configure appropriate limits in our OpenAI account. Could you please clarify: The approximate tokens per minute (TPM) consumed per active call when using the assistant The requests per minute (RPM) generated per call Any recommended TPM/RPM limits per concurrent call when using models like GPT-4.1 or GPT-5.2 in Vapi This will help us ensure our OpenAI account limits are configured high enough to avoid 429 rate limit issues when running campaigns with multiple concurrent calls. Thanks!

Hi, In terms of RPM, a typical call generates around 3–5 requests per minute. If your calls involve frequent interruptions, this can go up to 8–10 RPM per call. For TPM, short or simple calls generally consume around 5,000–10,000 tokens per minute, standard calls around 10,000–15,000, and longer or more complex calls can reach 15,000–25,000 TPM. If you're using structured outputs, expect to add roughly 50,000–70,000 TPM on top of those figures per call - it's a significant multiplier. As a rule of thumb, multiply your per-call estimates by your expected concurrency and add a 1.5x safety buffer. For production campaigns we'd recommend at least OpenAI Tier 2–3, and Tier 5 if you're running 100+ concurrent calls. These are ballpark estimates - actual usage will vary based on call length and complexity, so it's worth monitoring your OpenAI usage dashboard after go-live and adjusting limits from there. Let us know if you have any other questions! Best regards, Vapi Support