Deepgram Voice AI had 34 outages in the last 2 years totaling 1250h 57m of downtime — averaging 1.4 incidents per month.
There were 34 Deepgram Voice AI outages since July 24, 2025 totaling 1250h 57m of downtime. Each is summarised below — incident details, duration, and resolution information.
Some API requests may have experienced elevated error rates for approximately 2-3 minutes starting at 16:43 UTC on May 22, 2026. The issue has been resolved now.
Some API requests may have experienced increased latency and elevated error rates for approximately 30 minutes starting at 16:40 UTC on May 22, 2026. The issue has been identified and is being resolved.
Between 22:06 and 22:12 UTC, a subset of Voice Agent requests returned elevated 4xx errors following a recent deploy. The deploy has been rolled back and the service is operating normally.
We are currently investigating an increase in 5XX errors for non-English STT requests that are using redaction.
This issue has been resolved.
During the affected time period, some English batch speech-to-text requests intermittently returned 404 Not Found HTTP errors. This issue was limited to requests that met both of the following criteria: - Utilized English batch STT models (including both monolingual English and multilingual models that support English transcription) - Had Smart Format, Entity Detection, or Redaction features enabled.
This incident has been resolved.
We are currently observing elevated error rates from our downstream third-party provider, OpenAI specifically with the ChatGPT 4o model. Deepgram infrastructure, including our canaries, remains healthy.
This incident has been resolved.
This impacted STT fetch requests, callbacks, and voice Agent third-party providers. API calls that proxy to the console were also returning 503s intermittently.
We are currently observing elevated error rates from our downstream third-party provider, Anthropic specifically with the Claude Sonnet 4.5 model. Deepgram infrastructure, including our canaries, remains healthy.
This incident has been resolved.
We are currently observing elevated error rates from our downstream third-party provider, Google Gemini LLMs. Deepgram infrastructure, including our canaries, remains healthy; however, Voice Agent customers using bring-your-own (BYO) Gemini LLMs may experience degraded performance or intermittent failures on LLM think requests. We are monitoring the situation and tracking the provider’s status at https://aistudio.google.com/status. We will provide further updates as more information becomes available or if we observe broader impact.
This incident has been resolved.
We are currently seeing elevated error rates from OpenAI models. To avoid downtime, please define multiple LLM providers (https://developers.deepgram.com/docs/voice-agent-llm-models#using-multiple-llm-providers) in your Voice Agent configuration.
We are continuing to investigate this issue.
This incident has been resolved.
A networking incident was identified which resulted in failed callbacks on STT PreRecorded API requests and TTS Batch requests. The issue also affected Voice Agent API sessions with LLM and TTS providers. We have identified a fix and are monitoring traffic. The source of the issue is being investigated.
A network infrastructure fix was implemented at 18:15 UTC and the behavior was immediately resolved. We are continuing to monitor carefully.
This incident has been resolved.
There is an ongoing OpenAI outage that is affecting our Voice Agent. To avoid downtime, please define multiple LLM providers (https://developers.deepgram.com/docs/voice-agent-llm-models#using-multiple-llm-providers) in your Voice Agent configuration.
This incident has been resolved
There is an ongoing OpenAI outage that is affecting our Voice Agent. To avoid downtime, please define multiple LLM providers (https://developers.deepgram.com/docs/voice-agent-llm-models#using-multiple-llm-providers) in your Voice Agent configuration.
We are continuing to monitor this issue.
This incident has been resolved.
We are currently investigating an issue that results in increased Speech-to-Text result latency.
We are monitoring the issue.
This incident has been resolved.
A temporary issue during a Flux-related update caused intermittent SDK session errors. The issue was identified and resolved. 2:00 AM UTC - 13:45 UTC
We saw elevated connection errors from 19:12 to 19:19 UTC, and 20:10 to 20:31 UTC. We're no longer seeing spikes in errors, but we're continuing to monitor the situation.
We're observing elevated connection error rates at the moment, and we're actively investigating the issue.
We are continuing to investigate this issue.
We're currently seeing elevated response time, as well as increased 429 errors for batch STT. We're actively working on mitigating the issue.
We are continuing to investigate this issue.
A fix has been implemented and we are monitoring the results.
The implemented fixes have resulted in stable systems since 2026-02-03 17:35 UTC. We are continuing to monitor carefully. The issue primarily impacted STT batch services, with some impact on STT streaming. Specifically, response times and 429 errors increased for batch, and a percentage of connection errors occurred for both batch and streaming.
This incident has been resolved.
From 21:38 - 23:39 UTC on 2026-01-22, requests for Deepgram's Nova-3 multilingual STT model received degraded transcripts due to a faulty model release.
The issue has been identified and is being actively resolved.
A fix has been implemented and we are monitoring the results.
This incident has been resolved.
We're experiencing an elevated level of API errors and are currently looking into the issue.
We are continuing to investigate this issue.
We are still investigating the issue.
The issue has been identified and a fix is being implemented.
A fix has been implemented and we are monitoring the results.
We are continuing to monitor for any further issues.
We are continuing to monitor for any further issues.
This incident has been resolved.
We are investigating an issue causing elevated latency and connection timeouts to Deepgram services.
We have adjusted our networking and are seeing recovery.
This incident has been resolved.
We're experiencing an outage for our non-English batch speech-to-text requests utilizing redaction feature.
We've implemented a fix and are currently monitoring the status of the issue.
We are continuing to monitor for any further issues.
This issue has been resolved.
We are currently investigating increased latency over our streaming STT API and Voice Agent API.
We identified a potential source of increased latency and are currently monitoring a mitigation step.
The increased latency issue has significantly improved. We will continue to monitor the performance.
This issue has been resolved.
A service degradation and partial outage began at approximately 16:44 UTC.
We're seeing elevated error rates (500 Internal Server Error) with OpenAI models in the Voice Agent API. Our team is actively investigating and communicating with OpenAI to find a path to resolution, and a path to mitigation in the short term. Our metrics show degraded service for 4o-mini, 4.1-mini, 4.1 nano, 4.1, and possibly other OpenAI models. We are continuing to monitor, investigate, and work with OpenAI to resolve. As a workaround, we recommend switching to a different model or provider.
We are continuing to see elevated error rates with OpenAI models in the Voice Agent API workflow. We have reached out and communicated with OpenAI to help resolve.
We are still continuing to see increased errors using the OpenAI models in the Voice Agent API workflow. We have reached out and communicated with OpenAI to help resolve.
We are still continuing to see increased errors using the OpenAI models in the Voice Agent API workflow. We have reached out and communicated with OpenAI to help resolve.
As of 19:28 UTC OpenAI's LLMs are no longer returning 500 errors. We are still in conversation with OpenAI about what happened.
We're seeing elevated error rates with OpenAI models in the Voice Agent API. Our team is actively investigating and working with OpenAI to find a path to resolution, and a path to mitigation in the short term. Our metrics show degraded service for 4o-mini, 4.1-mini, 4.1 nano, 4.1, and possibly other OpenAI models. A service degradation began at approximately 23:00 UTC on October 9th. The service degraded further at 22:00 UTC on October 14th, and then again at 23:00 UTC on October 15th. We are continuing to monitor, investigate, and work with OpenAI to resolve. As a workaround, we recommend switching to a different model or provider.
As of 21:00 UTC on October 16th, OpenAI's LLMs are no longer returning 500 errors.