How Vannus Room routes prompts

The Vannus Room Pro+ elimination engine isn't a black box — it's a small set of published rules. We check them in order; the first match wins. You can audit every routing decision by typing /why in Room after any response. This is the manifesto applied to our own product: structural transparency.

Why Room, instead of subscribing direct?

If you already pay for ChatGPT Plus, Claude Pro, and Gemini Advanced separately, that's $60+/mo for three accounts — and you're still using one model at a time, manually tab-switching. Pro+ runs all five through a single workspace, routes each prompt to the best-suited model automatically, gives you /why auditability, and caps your total monthly compute at a known number. Two models you probably don't have direct subscriptions to (DeepSeek for code, Llama on Groq for speed) are included.

The transparency isn't a cost — it's the point. We tell you exactly which model runs and why, and you keep paying $59/mo regardless of how vendor pricing shifts under the hood.

Current rules (Pro+ tier)

Long-context

Prompts >~16k input tokens route to the long-context model. Gemini family preferred for ≥1M token windows.

Gemini 2.5 Pro / Flash

Code

Programming, debugging, code review, language-specific syntax questions.

DeepSeek V3 (preferred) → Claude Sonnet (fallback)

Creative

Long-form writing, ideation, copy, storytelling, voice work.

Claude Sonnet

Factual

Knowledge lookup, summarization, structured Q&A, citation-shaped answers.

Gemini 2.5 Pro

Reasoning

Multi-step logic, math, analysis, planning, strategic decomposition.

Claude Sonnet (preferred) → DeepSeek V3 (fallback)

—

Default

Everything that doesn't match a category above. Routes to the fastest, most cost-effective model available.

Gemini 2.5 Flash

How this works

One model per prompt. Unlike Pro (which fires two models side-by-side and lets you pick the winner), Pro+ uses these rules to pick a single best-fit model per prompt — keeping cost predictable while still adapting to what you're actually asking.

The rules are intentionally simple. No black-box ML classifier. No vendor preference. Just published, inspectable logic anyone can read. Phase 2 will replace the regex layer with an embedding classifier trained on real prompt data — but the categories and the model assignments stay public.

Models rotate. If a provider goes down or changes pricing in a way that breaks neutrality, we swap. Your subscription terms don't change. This page always reflects the current assignments.

The /why command. Inside Room, type /why after any Pro+ response and we'll explain exactly which rule fired, which model was selected, and what the preference order was. Zero compute cost — it's all metadata the routing engine already records.

Pro tier (side-by-side)

The Pro tier doesn't use this routing table — every Pro prompt fires two cost-efficient models in parallel, and you pick the winner. The "eliminated" answer is visibly dimmed. That's elimination made literal.

If you'd rather have Vannus pick for you, upgrade to Pro+.

Last updated by Vannus' deployment — rules are loaded live from /api/room/routing.