Replicate is a cloud platform for running and deploying machine learning models through a simple API. Developers use it to call image, video, audio, and language models without managing their own GPUs, and to push custom models with Cog so they run as scalable, on-demand endpoints. Because so many apps wire Replicate directly into their backends, an outage rarely looks like a single broken button. It looks like image generations that never return, webhooks that stop firing, or API calls that hang until they time out. The verdict and recent user reports at the top of this page tell you whether the problem is widespread right now or limited to your own setup.
Common reasons Replicate stops working
- API errors or timeouts: Calls to the prediction endpoint return 5xx errors, hang, or never resolve, which usually points to a platform-side incident.
- Long cold starts: A model that has been idle has to load onto a GPU first. Heavy traffic or capacity limits can stretch this into long waits that feel like a hang.
- Predictions stuck in queue: Jobs sit in “starting” or “processing” far longer than usual when GPU capacity is constrained or a region is degraded.
- Webhook delivery failures: Your prediction finishes but the completion webhook never arrives, leaving your app waiting on a result that already exists.
- Authentication and billing blocks: An expired or revoked API token, a spending limit, or a payment issue returns 401 or 402 errors that look like downtime.
- Rate limits and quota: Bursts of concurrent requests get throttled, returning 429 responses.
- A specific model is broken: One model version may fail or be removed while the rest of the platform is healthy.
How to fix Replicate when it is not working
- Check the verdict first. Read the live status and user reports at the top of this page to see whether others are affected before you start debugging your own code.
- Read the exact error. A 401 or 402 is an account problem, a 429 is rate limiting, and a 5xx or timeout points to the platform. Match your fix to the code.
- Verify your API token. Confirm the token is valid, not revoked, and present in your request headers. Rotate it if you suspect it leaked or expired.
- Check billing and limits. Make sure your payment method is current and you have not hit a spending or quota cap.
- Retry with backoff. For timeouts and 429s, retry after a short, increasing delay rather than hammering the endpoint, which only deepens throttling.
- Poll instead of relying on webhooks. If completion callbacks stop arriving, fetch the prediction by its ID to confirm whether it actually finished.
- Test a different model. If one model fails but others succeed, the issue is that model version, not Replicate as a whole.
- Allow for cold starts. If responses are slow but eventually return, the model may simply be booting. Keeping it warm can reduce this.
Is Replicate down, or is it just you?
Before assuming a platform outage, rule out local causes. A revoked token, a hit spending limit, a single broken model version, or a network or firewall block on your end can all mimic downtime. Try the API from a different network or machine, and test a known-good model. If your calls fail everywhere and the reports above show a spike, the problem is almost certainly on Replicate’s side. If you are the only one reporting trouble, focus on your token, billing, and code. For a structured walkthrough, see our guide on what to do when an AI tool is down.
How we check Replicate
We combine two independent signals so you are not relying on a single source. First, we read Replicate’s official status page to catch incidents the company has acknowledged. Second, we run our own reachability check from outside their network, hitting the service directly to confirm whether it actually responds for real users. When those two signals disagree, that gap is often the earliest sign of trouble. You can read the full details in our methodology.
Alternatives to Replicate during an outage
If Replicate is down and you need to keep shipping, several peer platforms can host or serve models while you wait. Hugging Face offers a huge model catalog and inference endpoints. Together AI runs many open models behind a fast API. Groq is built for very low-latency inference on supported models. Switching providers usually means changing endpoints and adapting request formats, so it is best as a planned fallback rather than a scramble. Before you cut over, confirm the backup is healthy on its own status page, since AI infrastructure outages sometimes overlap and you do not want to move onto a service that is also struggling.
Cold starts and other Replicate quirks
Replicate runs models as predictions you create through its API or web site. The single most common source of confusion is not an outage at all: cold starts. When a model has been idle, the first request has to boot the container and load weights, which can take many seconds or longer before any output appears. A prediction stuck in a starting or booting state is usually a cold boot, not a failure, and it typically resolves on its own once the model is warm.
Genuine problems show up as predictions that move to a failed status with an error, 429 responses when you exceed rate limits, 402 or billing related rejections when a payment method or balance is the issue, and 5xx errors during platform trouble. The web site and the prediction API are separate surfaces, so you can browse models or sign in while predictions misbehave, or the reverse. Because anyone can publish models, a failure can also be specific to one model version rather than Replicate as a whole.