Academy
4. Specialists: Reasoning, Code, and Live Web
Know when to reach for a reasoning model, a code model, or a live-web model instead of a general-purpose one.
OUPI's specialist family contains three kinds of models, each built for a job that general-purpose models handle less well.
• Reasoning models (e.g. GPT-5.6 Sol, Magistral, DeepSeek R1) spend extra time thinking before they answer. They chain intermediate steps internally, which makes them stronger on multi-step calculations, constraint-heavy decisions, multi-criteria comparisons, and tricky debugging. The trade-off: they are slower and cost more credits per exchange.
• Code models (e.g. Codestral, Kimi K2.7 Code) are optimised for generating, completing, and refactoring code. They are strong and economical for pure coding tasks. For building a full application, OUPI also offers CodeBuilder, an agent that plans, writes, runs, and fixes code around the model.
• Live-web models (Perplexity) can fetch current information from the internet. A regular model only knows what existed up to its training cutoff, so for today's news, live prices, or recent regulations, a live-web model — or turning on web search — is the right call.
When to reach for a reasoning model
Use a reasoning specialist when the answer depends on chaining several logical steps: a financial calculation with intermediate results, a comparison across many weighted criteria, a planning problem with constraints, or isolating a subtle bug in code. If a balanced model gives a shallow or wrong answer on such a task, switching to a reasoning model often fixes it.
Do not use a reasoning model for rewriting, summarising, or simple Q&A — it will be slower and more expensive without being noticeably better. The rule of thumb: if the task needs thinking, pick a reasoning model; if it needs writing, a balanced model is enough.
When to reach for a code model
Frontier and balanced models already write good code, so a code specialist is not always necessary. Reach for Codestral or Kimi Code when:
• You need high-volume code generation or completion and want to keep costs down. • The task is purely technical — no mixed prose-and-code output. • You are iterating quickly on snippets and value speed.
For end-to-end application building (planning, writing, running, debugging), use CodeBuilder instead of chatting with a code model directly — it wraps the model in an agent workflow that handles the full cycle.
When to reach for a live-web model
Every non-web model has a training cutoff: it literally does not know what happened after that date. Choose a live-web model (Perplexity) or enable web search when your question involves:
• Current events, recent statistics, or today's prices. • A regulation or API that may have changed since the model was trained. • Fact-checking a claim against the latest sources.
For images, PDFs, and audio, you do not need a specialist — any recent multimodal model (Claude, GPT, Gemini families) can read them. Image generation, video, and speech have their own engines in Studio and the media tools.
Specialists cost more credits than fast models and are often slower. A good habit: start with a balanced model, read the answer critically, and escalate to a specialist only for the turn that actually needs it. You can switch models mid-conversation without losing history, so escalation is free of friction.
Open the OUPI chat interface. Ask a multi-step maths or logic question using a balanced model first. Then switch to a reasoning specialist (e.g. GPT-5.6 Sol or Magistral) on the same conversation and compare the depth of the two answers. Notice how the history is preserved after the switch.
Three specialist types exist alongside the general families: • Reasoning models — for multi-step logic, calculations, and complex debugging. • Code models — economical for pure code generation; use CodeBuilder for full apps. • Live-web models — for anything that needs information newer than a model's training cutoff.
Start with a balanced model by default. Escalate to a specialist only when the task demands it, and remember you can switch mid-conversation without losing context.