Academy
1. Why OUPI Offers Many Models — and Four Families to Know
Understand the landscape of model families (frontier, balanced, fast, specialist) so the automatic selector's choices make sense to you.
OUPI connects you to models from multiple providers — Anthropic (Claude), OpenAI (GPT), Google (Gemini), Mistral, DeepSeek, Moonshot (Kimi), xAI (Grok), and Perplexity for live web data. No single model excels at everything, and relying on one provider is a risk. That's why OUPI gives you a whole landscape of options and can route each message to the right one. To navigate this landscape, you only need to remember four families:
• Frontier — the most capable models for hard reasoning, long documents, and complex agent work (e.g. Claude Opus 5, GPT-6 Astra, Gemini 3 Pro). • Balanced — excellent quality at lower cost; the right default for most professional work (e.g. Claude Sonnet 5, GPT-5.6 Terra, Mistral Medium, Gemini 3.8 Flash). • Fast — for short, high-volume, or simple tasks (e.g. Claude Haiku 4.5, GPT-5.6 Luna, Mistral Small, Gemini Flash Lite). • Specialist — reasoning models that think longer (GPT-5.6 Sol, Magistral, DeepSeek), code models (Codestral, Kimi Code), and live-web models (Perplexity).
Why four families instead of a flat list? Because provider names and model versions change every few months, but the families stay stable. When a model you knew disappears, look for its family — frontier, balanced, fast, or specialist — and pick the current member. This mental map is also exactly what OUPI's automatic selector uses: it classifies your task, then picks from the right family based on complexity, cost, and reliability.
Five criteria help you (and the automatic selector) decide which family fits a task:
- Difficulty — How much reasoning is needed? A rewrite is easy; a multi-step legal analysis is hard.
- Length to read — How big is the input? Frontier and balanced models handle up to ~1 million tokens; lighter models read far less.
- Modality — Text only, or images/PDFs/audio too? Recent Claude, GPT, and Gemini families are multimodal.
- Data sensitivity — Does it matter where data is processed? Mistral is French/EU-hosted; admins can restrict models per team.
- Budget — Frontier models cost roughly 10–30× more per token than fast ones. A balanced model for bulk work plus a frontier model for the hard part is the usual sweet spot.
When you don't pick a model yourself, OUPI's automatic selection reads your message — the kind of task, its complexity, whether files or tools are involved — and scores every model your plan allows on fitness, cost, and recent reliability. A routine request gets a fast model (quicker, cheaper); a complex one escalates to balanced or frontier automatically. A badge under each answer tells you which model was used, so the choice is never hidden.
Start with automatic selection or a balanced model. Move up to frontier only when the task is genuinely hard — long multi-document synthesis, complex code, or tricky reasoning. Move down to fast for repetitive, short, or low-stakes work. This keeps quality high and costs low.
Need a reasoning specialist? Use one when the answer depends on chaining several steps — calculations, multi-criteria comparisons, constraint-based decisions. They think longer and cost more, so they're wasted on simple rewrites or summaries. The automatic selector handles this trade-off per message if you let it.
Open the OUPI chat and send two messages without picking a model: one simple ("Rewrite this sentence more formally") and one complex ("Compare the pros and cons of three options and recommend one"). Check the badge under each answer — you should see a lighter model for the first and a stronger one for the second.
OUPI's models fall into four families: frontier (hardest tasks), balanced (daily default), fast (simple/high-volume), and specialist (reasoning, code, live web). Five criteria guide the choice: difficulty, input length, modality, data sensitivity, and budget. The automatic selector applies these criteria per message so you get the right model without thinking about it — but you can always override manually, and the badge always shows what was chosen.