It won't fail loudly, tool calls just come back unparsed instead of structured. Match the parser to your model family (for example, Qwen3 chat models use qwen3_xml, gpt-oss uses openai).

You pay for the session length you set, in the environment's accepted fee token. The full cost is locked into an escrow contract on Base and claimed by the node upfront, before your session starts. Nothing more is charged unless you prolong the session.

Yes. Prolong session adds runtime, priced against the resources the service is currently holding. This isn't available once a service has expired, at that point a new launch is the only way forward.

A model launch gives you an API, you call it programmatically. A Service gives you an app: its own web UI in a browser tab, running on the node's hardware. Services also ship with no models pre-loaded; you load your own checkpoints, LoRAs, or files from inside the app once it's running.

The container is torn down and the endpoint stops answering. Prolong before the countdown hits zero if you want to keep it running.

No. The direct route skips the resource step entirely, the service's own recommended resources size the booking for you, so you set session length, click a price, and go straight to payment. Advanced setup is there if you want to size the hardware yourself instead.