3. Five Criteria for Choosing

Every time you pick a model in OUPI, five criteria should guide you:

1. Difficulty — How much reasoning does the task need? A simple rewrite needs less power than a multi-step legal analysis.
2. Context length — How much must the model read at once (your message, conversation history, attached files)? Recent frontier and balanced models handle up to roughly a million tokens; lighter models handle far less.
3. Modality — Do you need the model to process images, PDFs, or audio, or just text? Not every model is multimodal.
4. Data sensitivity — Does the content require a European provider, or has your admin restricted certain models? OUPI never uses your content to train any model.
5. Budget — Frontier models can cost ten to thirty times more per token than fast ones. A long frontier conversation is the most expensive thing you can do in chat.

Speed is a bonus, not a sixth criterion: a fast wrong answer costs more than a slow right one.

Criterion 1 — Difficulty. Start with a balanced model; it handles writing, summaries, analysis, and most professional work very well. Move up to a frontier model only when the task is genuinely hard — complex reasoning, long multi-document synthesis, or tricky code — or when a balanced model has visibly struggled. Move down to a fast model for repetitive, short, or low-stakes work.

If the task requires chaining several logical steps (calculations with intermediate results, multi-criteria comparisons, constraint-based decisions), consider a reasoning specialist. These models think longer before answering, making them slower and more expensive — but they are wasted on simple rewriting or summarizing.

Criterion 2 — Context length. The context window caps how much the model can read at once. If you need to process a long contract, a large spreadsheet, or a conversation with extensive history, choose a model whose window fits the material.

When your documents exceed the window — or when you want to avoid pasting confidential text into prompts — use a knowledge base instead. OUPI retrieves only the relevant passages, so the model reads what matters without hitting the limit.

Criterion 3 — Modality. Images and PDFs can be read by any recent multimodal model (Claude, GPT, Gemini families). For live, up-to-date information, use a live-web model like Perplexity or turn on web search, because regular models only know up to their training cutoff. Image generation, video, transcription, and speech use dedicated engines in Studio, not chat models.

Criterion 4 — Data sensitivity. When handling confidential or regulated data, prefer European providers — Mistral is French and hosted in Europe — or whichever provider your organization has approved. Your administrator can restrict which models your team may use, giving you a curated list that already meets policy.

Remember: OUPI never uses your content to train any model. For extra caution with sensitive material, use a knowledge base so your documents stay inside OUPI rather than being pasted into prompts.

Criterion 5 — Budget. Credits are proportional to the tokens exchanged multiplied by the model's weight. Frontier models consume roughly ten to thirty times more per token than fast ones. The cost-effective pattern: use a balanced model for the bulk of your work and escalate to a frontier model only for the hard parts. Your balance and each exchange show exactly what was consumed.