How does Jarvis handle multiple AI models?

Jarvis (getjarvis.eu) routes per task across frontier models from Anthropic, OpenAI, and Google (voice uses OpenAI realtime) automatically based on task type. The router considers: prompt complexity, expected output length, visual context (multimodal prompts go to whichever model is strongest for the task), latency budget, and your tier. Default routing: a frontier model for nuanced reasoning and tone-matching, a faster model for fast triage and structured output, with longer-context jobs sent to a frontier model. You can override routing per prompt or set defaults in Settings → Models. The router runs server-side in the Jarvis backend and all calls go to provider enterprise API tiers — your prompts are not retained for training.

No single model wins at every task. Jarvis routes per task across frontier models from Anthropic, OpenAI, and Google (voice uses OpenAI realtime). (1) a frontier model is the strongest at nuanced reasoning, tone-matching, code review, and instruction following — particularly for long, complex prompts. (2) a faster model is fast and excellent at quick triage and short-form, structured output. Routing means you get the best model per task without manually picking each time. The router is conservative — for ambiguous prompts it defaults to a frontier model because it's the most consistent across task types. Users who want manual control can pin a model in Settings → Models.

The Jarvis router is a small heuristic classifier (not an LLM) that runs in the Jarvis backend. Inputs: (1) Prompt length and complexity (token count, sentence structure, presence of code blocks). (2) Connector data attached (a 500-message Slack thread → Gemini for the long context). (3) Visual context (multimodal prompts route to whichever vision model is best for the task; a frontier model by default in 2026). (4) Latency budget (quick triage prompts default to a faster model). (5) Your tier (free tier prefers cost-efficient routing, Pro and Unlimited prioritize Opus). (6) Your stated preference if set in Settings → Models. The router runs in <10ms — invisible to the user. You can see which model handled each prompt in the floating bar's response metadata.

AI models