DeepSeek credits users with tokens as compensation after multi-hour outage

DeepSeek lost its web chat and API for several hours on October 15. Around 8 a.m. the service began responding with "server temporarily unavailable," and the status page showed three components affected at once: the chat, the V4 Pro API and the V4.1 Flash API. The company later credited users with tokens as compensation. For the Chinese market this is a notable step: model providers usually stop at an apology and a postmortem.

How the outage unfolded

The first complaints arrived at about 8 a.m. The web version and the API stopped responding almost simultaneously, and for many users at once, so the problem looked widespread from the start rather than local. The official status page classified the incident as a partial outage. In other words, some requests could get through, but stable operation could not be counted on.

By 9 a.m. DeepSeek had updated its announcement: the cause had been identified, and work on restoring service was under way. In the portion of the Jiqizhixin article available to us, the text of the announcement cuts off, so we do not name the details of the root cause (network, inference cluster, deployment, load): the company has not disclosed them publicly. The headline of the original refers to downtime of several hours.

One important detail: it was not only the consumer chat that went down but also two paid API lines. That turns the incident from an inconvenience for app users into a problem for developers who have built products on DeepSeek.

Why compensate in tokens

The choice of currency is logical here. DeepSeek's API is billed per token, so crediting tokens fits directly into the existing billing model: no refunds, invoices or accounting adjustments are needed. It is also a cheap gesture for the company itself. The cost of the issued tokens is the cost of compute, not the retail price, and in the short term they require no cash.

There is a second effect. Tokens tie the customer to the platform: someone who received credit is more likely to spend it on DeepSeek than to go to a competitor. This is our interpretation, not a statement from the company. The size of the compensation and the terms for receiving it are not given in the portion of the article available to us, so the generosity of the gesture cannot be judged yet.

Still, the fact itself is telling. Companies whose models have become part of working processes increasingly have to behave like cloud providers: with a public status page, root-cause analyses and compensation for downtime.

What is known so far

• Outage start: around 8:00 a.m. on October 15.
• Affected: web chat, V4 Pro API, V4.1 Flash API.
• Status page classification: partial outage.
• Around 9:00 a.m.: DeepSeek said the cause had been found.
• Duration: several hours, according to the Jiqizhixin headline.
• Compensation: tokens; the size and terms are not named in the available materials.

What teams that work through the API should do

First: check your balance and credit history in your account dashboard. If the compensation is automatic, it should appear there; if you need to file a request, it is better not to put it off.

Second: look at how your product handled the morning. If requests simply failed with an error, now is the time to add retries with exponential backoff and timeouts that do not freeze the user interface.

Third: set up a fallback route. A backup model from another provider, or a local open model for critical scenarios, costs little compared with several hours of downtime. Any request router that switches automatically on an error code will do the job.

Finally, subscribe to the status page and set up alerts on your own error metrics. Learning about a problem from customer complaints is slower than learning about it from monitoring, and in DeepSeek's case the window between the outage and the first official message was about an hour.