Models and keys

Without a key of your own, runs use the plan's model credits on the Gemini models, which Pocket serves from Google Cloud (Mumbai, with Google's global endpoint as backup when the region is busy): Gemini 2.5 Flash, Flash-Lite and Pro, priced at Google's list price plus 15%, metered per request from the usage Google reports, with nothing sold through a reseller. Pro includes $2 of credits a month, Free $0.25; top-ups add more. Under Account → Models & keys, add your own key for a provider and runs on it use your key instead, with no credits taken and no markup; that is also the only way to run models from other vendors (OpenAI, Anthropic, OpenRouter, Fireworks, Sarvam and twenty-some more, or any OpenAI-compatible endpoint). The catalogue shows what each key unlocks, and the model picker says which route a model takes.

• Per project default and per conversation: pick the model in the top-bar chip; the choice becomes the project's default for the next conversation.
• Calls use a fast model separate from the chat model (Gemini 2.5 Flash unless you pick another), so the first word comes back in about a second.
• Provider health is checked hourly. You get a push and a WhatsApp when a key's credits run low or a provider is down, so bots do not stop mid-call; speech providers fall over to the next key during a call.
• On a team, owners decide whether members may bring their own keys or everyone uses the team's.

Keys are encrypted with a master key the API alone holds and are never returned to the app; you see a hint, not the key. What each model costs.