Deepgram

The upstream ISP failure has been mitigated by rerouting all traffic around that provider.

Cogent is experiencing a backbone networking issue in LA, due to two fiber cuts. This is likely impacting a large majority of traffic flowing to the east-coast, we're looking into possible mitigations and Cogent has engineers actively working on it. https://ecogent.cogentco.com/network-status

Resolved Between 07:13 and 09:00 UTC, flux-general-multi Speech-to-Text requests on api.eu.deepgram.com failed shortly after the WebSocket connection was established, returning an internal server error on the first audio and closing the connection. Affected clients saw abnormal closures and reconnect loops. A low rate of residual failures continued until 11:30 UTC. This has been fully investigated, mitigated and resolved, service has fully recovered.

Between 23:18 and 23:45 UTC, a portion of streaming STT connection requests failed with 5XX errors and timeouts. While rolling out a new data center, we detected streaming errors and immediately removed it from rotation. Traffic to the new site had been intentionally limited to a small percentage while we verified the rollout, which limited the number of affected requests. This is now fully resolved.

This incident has been resolved.

A fix has been implemented and we are monitoring the results.

We're seeing elevated error rate for authentication services. We are currently investigating the issue

We've resolved the issue and the service is back to normal.

We're seeing elevated 503 error for v1/auth/grant endpoint.

This incident has been resolved.

An ISP network failure resulted in poor performance along one network path, resulting in degraded performance that primarily affected STT traffic. All network traffic is now routed through a high quality network path and services have recovered. We are actively monitoring.

We are currently investigating an issue where batch and streaming STT requests have degraded performance.

This incident has been resolved.

A fix has been implemented and traffic has returned to normal. We are monitoring the results.

We are continuing to investigate an issue. STT batch requests may have returned 5xx errors, and some STT websockets may have experienced high latency, starting at 16:03 UTC.

We are continuing to monitor for any further issues.

A fix has been implemented and we are monitoring the results.

We are continuing to investigate this issue.

We are currently investigating an issue where some batch requests are returning 400 errors due to remote content download errors. Retrying the request one or more times often succeeds.

This incident has been resolved.

Between approximately 10:55 and 12:00 UTC on July 24, an upstream infrastructure issue in AWS's US-WEST-2 region caused intermittent connectivity problems that may have affected Deepgram services and customer infrastructure — including elevated errors, connection timeouts, and failures fetching remotely hosted media for some requests. The underlying cause was an AWS networking issue affecting routing to the US-WEST-2 region. AWS applied mitigations and fully restored connectivity by ~12:00 UTC. We have confirmed recovery and are continuing to monitor. Details on the underlying AWS event are available on the AWS Health status page: https://health.aws.amazon.com/health/status?eventID=arn:aws:health:us-west-2::event/MULTIPLE_SERVICES/AWS_MULTIPLE_SERVICES_OPERATIONAL_ISSUE/AWS_MULTIPLE_SERVICES_OPERATIONAL_ISSUE_83A36_24C89B493A6

This incident has been resolved.

Starting at 13:45 PM UTC, some Voice Agent sessions using Google/Gemini without a pinned model version received failed responses due to a routing issue introduced in this morning's deploy. Two Gemini models were unaffected; all other providers and products were unaffected. The issue was resolved as of 3:20 PM UTC.

This incident has been resolved.

Flux STT experienced elevated WebSocket streaming errors between 15:56 UTC and 18:20 UTC.

This incident has been resolved.

Continuing to monitor; root cause confirmed as the Google API key change.

We have identified the cause of elevated error rates affecting Voice Agent requests that use Google Gemini models. A recent change on Google's side to how API keys operate is causing requests to fail, and our team is rolling out updated credentials to resolve it. Until the update is complete, customers may see intermittent failures for Gemini-model requests — concentrated in the first half of each hour, with most requests succeeding in the second half. Other Voice Agent models are unaffected. We are actively working on the fix and will update when resolved.

We are currently investigating elevated error rates affecting Voice Agent requests that use Google Gemini models. During this period, some requests to Google models (including gemini-2.5-flash-lite) may fail or time out. Other Voice Agent functionality and non-Google models are not affected. Our team is actively working to identify the cause and restore full service. We will provide an update shortly. 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.

An upstream ISP issue may have caused intermittent WebSocket connection failures, handshake timeouts, and degraded streaming connectivity. Incident time: 01:10 - 01:52 UTC

The issue was resolved at 19:47 UTC.

Between the hours of 13:35 UTC - 19:47 UTC we noticed an increase in 400 errors returned on PreRecorded STT requests where the payload was a URL. They were transient errors, retrying the request was the recommended course of action and would typically result in a successful transcription.

Anthropic has resolved this incident. Incident end time: Jun 23, 2026, 16:44 UTC

We are currently observing elevated error rates from our downstream third-party provider, Anthropic. Deepgram infrastructure, including our canaries, remains healthy. Incident start time: Jun 23, 2026, 14:19 UTC (Anthropic status updates: https://status.claude.com)

Nova-3 and Nova-2 streaming STT experienced elevated websocket connection errors and websocket disconnections between 13:00 - 16:00 UTC.

This incident has been resolved. A degradation affecting a subset of streaming Speech-to-Text (Flux) sessions, where connections terminated before returning a transcript, has been mitigated and service has been operating normally since. We're continuing to monitor and apologize for the disruption.

We identified a degradation affecting a subset of Flux streaming Speech-to-Text sessions, where connections terminated before returning a transcript. We've isolated it to specific degraded hosts and are mitigating.

Some streaming Speech-to-Text connections on the Flux model established successfully but then terminated a few seconds into the session due to an internal error, before returning a final transcript. Applications relying on final transcripts (e.g. IVR/voice-agent flows) would have seen sessions open with no transcript returned.

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.

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.

This issue has been resolved.

We are currently investigating an increase in 5XX errors for non-English STT requests that are using redaction.

This incident 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.

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.

A small number of customers experienced disruptions to STT fetch requests, callbacks, and third-party providers within Voice Agent, as well as 503 errors on API calls for usage lookups and token creation. The majority of customers were unaffected.

This incident has been resolved.

We are continuing to investigate this issue.

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.

This incident has been resolved.

A network infrastructure fix was implemented at 18:15 UTC and the behavior was immediately resolved. We are continuing to monitor carefully.

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.

Between 15:16 and 15:32 UTC on March 25, 2026, we noticed an increased rate of API requests returning errors, including 403, 404, and 504 status codes. The issue has been identified and resolved.

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.

We are continuing to monitor this issue.

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.

This incident has been resolved.

We are monitoring the issue.

We are currently investigating an issue that results in increased Speech-to-Text result latency.

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

The upstream Anthropic incident affecting Claude models has been resolved and error rates have returned to normal. Voice Agent functionality should now be operating as expected.

There is an ongoing Anthropic 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. The incident started at Feb 17, 2026 - 20:15 UTC

As of 10:00 UTC, the issues from OpenAI models is being 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, including at least one non-OpenAI think provider. The incident started at 14:40 UTC.

This incident has been resolved.

We are continuing to investigate this issue.

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, including at least one non-OpenAI think provider. The incident started at 16:30 UTC.

The 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, including at least one non-OpenAI think provider. The incident started at 17:47 UTC

This incident has been resolved.

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.

A fix has been implemented and we are monitoring the results.

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.

We're observing elevated connection error rates at the moment, and we're actively investigating the issue.

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.

The downstream provider indicated the issue was resolved. There was an outage between 16:20 UTC / 8:20 PT and 16:55 UTC / 8:55 PT

We are currently seeing elevated error rates from Anthropic 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, including at least one non-Anthropic think provider. The incident started at 16:34 UTC

The incident has been resolved. The downstream provider's status page indicates the issues were resolved at 9:56 PT / 17:56 UTC

We are continuing to monitor for any further issues with the downstream provider.

Failures have reduced since 15:55UTC, we are monitoring the downstream provider's status page.

We are currently seeing elevated error rates from Anthropic 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, including at least one non-Anthropic think provider.

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.

This incident has been resolved.

A fix has been implemented and we are monitoring the results.

The issue has been identified and is being actively resolved.

Informational (No Customer Action Required) Deepgram identified that our internal development environment was affected as part of the industry-wide Shai-Hulud 2.0 NPM supply chain attack. The attack used compromised NPM packages to inject malicious CI/CD workflows and attempt to exfiltrate internal development credentials. Our investigation confirms: - No access to customer data or databases - No impact to production API infrastructure or service availability - No modification of published Deepgram SDKs or packages - No effect on customer authentication or API keys Timeline (UTC): - 00:41, Nov 26: Webhook alert on an internal GitHub repo; engineers link activity to Shai-Hulud 2.0, disable malicious workflows, and rotate exposed credentials. - 21:45, Nov 26: Additional commits via a compromised GitHub App publish some internal materials (no customer or production impact); the account is removed, GitHub Actions disabled, and the org moved into a locked-down state. We see no further signs of compromise and are gradually restoring normal operations while tightening SDLC and CI/CD controls. Customer action: None required. Questions: security@deepgram.com • support@deepgram.com This incident affected internal development infrastructure only; customer-facing APIs and services were not impacted.

This incident has been resolved.

We are continuing to monitor for any further issues.

We are continuing to monitor for any further issues.

A fix has been implemented and we are monitoring the results.

The issue has been identified and a fix is being implemented.

We are still investigating the issue.

We are continuing to investigate this issue.

We're experiencing an elevated level of API errors and are currently looking into the issue.

This incident has been resolved.

We have adjusted our networking and are seeing recovery.

We are investigating an issue causing elevated latency and connection timeouts to Deepgram services.

This issue has been resolved.

We are continuing to monitor for any further issues.

We've implemented a fix and are currently monitoring the status of the issue.

We're experiencing an outage for our non-English batch speech-to-text requests utilizing redaction feature.

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 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.

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'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.

A service degradation and partial outage began at approximately 16:44 UTC.

This issue has been resolved.

The increased latency issue has significantly improved. We will continue to monitor the performance.

We identified a potential source of increased latency and are currently monitoring a mitigation step.

We are currently investigating increased latency over our streaming STT API and Voice Agent API.

As of 21:00 UTC on October 16th, OpenAI's LLMs are no longer returning 500 errors.

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.

Start: 16:15:00 UTC End: 17:18:00 UTC Streaming connectivity issues from some AWS regions may have caused Websocket failures errors and degraded connectivity

This incident has been resolved.

The issue has been mitigated after rerouting traffic to avoid upstream network provider failures. Networking errors may have occurred intermittently between 12:08 am UTC and 1:51 am UTC.

We've identified an issue with an upstream network provider across various US regions, and are working to shift traffic to alternative routes.