Timeweb Cloud Adds Video Models, Ruble Limits and Fallback Models to AI Gateway

An AI bill usually grows quietly. Someone connected an expensive model to a test script, someone forgot to stop a loop, and at the end of the month it turns out the quarter's budget burned through in two weeks. Timeweb Cloud's October digest is first and foremost the provider's own product notes, and should be read with that in mind. But it contains a set of features that hit exactly this pain point.

One key for text, images and video

AI Gateway is a single entry point to models from different vendors. In October, video was added to it. The provider names Veo 3.1 from Google (a clip with sound right away), Kling from Kuaishou (a multi-shot scene with one character), Seedance from ByteDance, Gen-4.5 from Runway for animating product photos, Hailuo from MiniMax for vertical videos and Wan from Alibaba for long videos. For images, Seedream 5 and Gen-4 Image were added.

The scheme is simple. You send a prompt and, if needed, an image, and after a few minutes you pick up the finished clip. Images arrive immediately, up to four per request. To switch models, it is enough to change its name in the request. Video is billed per second of finished clip.

Imagine a marketplace team that needs to bring product cards to life. Before, that meant six different SDKs, six keys and six billing systems. Now there is one integration code, and comparing models comes down to changing a line in the request. Besides video, 20 more models appeared in the catalog: text models, embeddings, rerankers, speech recognition and synthesis. Each has a card with the price, a calculator, context size and information on where and by whom data is processed.

How to avoid a surprise in your bill

The main practical part is spending control:

• The "Activity" dashboard gathers spend, requests and tokens across all keys and agents. You can see who spends the most and which requests could be moved to a cheaper model.
• Request logs show the model, tokens, cost and status for each call.
• Daily limits in rubles are set per key or agent, plus separate limits per model. At 80% of spend a notification arrives, and at 100% new requests stop going through.
• Batch API accepts up to 50,000 requests at a time and returns answers as a file within 24 hours. Tokens cost half as much.
• Fallback models: up to two backups can be attached to a key. If the primary does not respond, the request goes to the next one, and you pay only for the one that responded.

A real-life example: overnight labeling of 40,000 reviews does not need an instant answer, so it makes sense to send it through Batch and save half. Meanwhile the support chat stays on a fast primary model with two backups, so that one vendor's outage does not take down the product.

What else is in the digest

AI agents now have long-term memory: they save facts from conversations, and the contents are visible in the "Memory" tab, where they can be edited or deleted. App Platform now spins up a separate environment for each pull request and deletes it after closing. Kubernetes gained private clusters and a connector that itself issues and renews the AI Gateway key as a secret. Dedicated servers received 47 GPU configurations, from RTX 4090 to H100 and H200, as well as backups to your own S3 with deletion protection. The rest, from domains to panel navigation, is mostly of interest to existing customers.

Limitations and takeaway

Note three points. First, this is the vendor's self-description: the model list, prices and availability should be checked in the catalog, not in the digest. Second, the service is oriented toward the Russian market, with settlements in rubles and local infrastructure, and will not suit every team. Third, a hard limit at which requests simply stop going through protects the budget but can take production down. Set it with a margin and be sure to configure the 80% notification.

Practical takeaway: if you call several models through different keys, look at a gateway as a control layer. Unified logs, limits, batch processing and fallbacks are useful regardless of which provider you get them from.