How does Jarvis route between Claude, GPT, and Gemini?

A small heuristic classifier runs in the Jarvis (getjarvis.eu) backend on every prompt and picks the best model. Inputs: prompt length, complexity, presence of visual context (multimodal), expected output length, connector data attached, latency budget, your tier, and your stated preferences. Default rules: a frontier model for nuanced reasoning and tone-matching, a faster model for fast triage and structured output, a long-context model for very long context (>150k tokens) and multimodal-heavy prompts. Override per-prompt with /opus, /gpt, or /gemini prefixes. Set defaults in Settings → Models. There is no bring-your-own-key mode — the router always calls Jarvis’s own provider accounts, never a tenant of yours. The router runs server-side in under 10ms — invisible to the user. All providers' enterprise API tiers are used, so prompts are not retained for training.

Five rule categories. (1) Token count: prompts > 150k tokens (including connector data) route to the model with the largest context window. 50-150k routes to a long-context model for high-accuracy recall over very large inputs. < 50k uses task-type routing. (2) Task type: nuanced reasoning / drafting / code review → a frontier reasoning model. Fast triage / structured output / function calling → a fast model. Long-context synthesis → a long-context model. (3) Visual context: multimodal prompts (screenshot, image upload) route to a frontier vision model by default; fast triage uses the faster model. (4) Latency budget: "quick" mode prompts (typed with /q prefix) prefer a faster model. (5) User preferences: Settings → Models lets you pin a default per category. All rules combine in a small weighted decision tree.

Three ways to override the router. (1) Per-prompt prefix: start your prompt with /opus to force the strongest available model or /flash to force a faster one. Useful when you have a strong preference for one prompt. Example: "/opus draft a careful reply to the legal email about the IP question." (2) Settings → Models defaults: set a default model per category (Drafting, Summarization, Code, Q&A). The router still considers token count for long context but otherwise honors your preference. (3) Pin a single model: in Settings → Models, pin "always use a frontier model." Useful for users who specifically want Anthropic's model voice. None of these modes involve your own API key: model access is managed for you, so the router always calls Jarvis’s provider accounts.

AI models