Two AI tiers: Frontier and Workhorse

Runwita uses AI for two very different shapes of work, and they have very different cost profiles. Rather than ask one model to do both well, the app splits them into two independently configurable tiers.

​

The two tiers at a glance

You configure both in Settings → Models, with their own provider, API key, and model dropdown. They’re independent, you can run Claude on one tier and OpenAI on the other if you want.

​

Frontier (the intelligence layer)

The Frontier tier handles work that needs to read the entire journey: every engagement, every topic, every decision, and synthesise something coherent across them. This is the kind of work where output quality compounds with model capability. What runs on Frontier:

• Deal stage detection. Reads the journey’s history and decides where in the customer lifecycle this is (Discovery, Qualification, Build, Go-live, Renewal, Churn Risk).
• Stakeholder analysis. Identifies the influence map across attendees over time.
• Sentiment analysis. How is this relationship trending, mood-wise.
• Deal health. A composite score with reasoning.
• Meeting brief. Pre-meeting prep: what was the last conversation, what’s open, what to land in this one.
• Objection detection. What concerns has this stakeholder raised that haven’t been addressed.
• Commitment gap detection. What did we say we’d do, and haven’t.
• Stale journey flagging. Has this gone quiet in a way that should worry us.
• Executive summary. A board-ready paragraph distilling the journey.

These features are explicit, you trigger them from the journey page (mostly via “Run all intelligence”). They’re not background magic; you choose when to spend the tokens. The default is Claude Haiku 4.5 for cost reasons, but Frontier is the place where switching to Opus or gpt-5 actually pays off. The journey context is large, the synthesis is hard, and the output goes onto persistent objects you’ll read again. Use the better model when the budget allows.

​

Workhorse (extraction and chat)

The Workhorse tier handles the high-frequency, mostly mechanical work:

• Meeting extraction. The big one. Every transcript, every set of notes, every email goes through this. Title, date, summary, sections, decisions, actions, attendees.
• Journey matching. When you save an engagement, picking which journey it belongs to.
• Topic matching. When you save an engagement, deciding which topics each section belongs to (or whether to create new ones).
• Chat. The chatbot UI for asking questions about a journey.

Workhorse runs constantly, on every save. Cost adds up. The default is gpt-5.4-nano (cheap and fast), but it’s also where the smaller models are weakest at the harder sub-tasks (topic matching especially). If you find your topics are over-fragmenting or your journey matches are flaky, upgrade Workhorse to Claude Haiku 4.5 or gpt-5.

​

When to upgrade which tier

A few rules of thumb: If you’re cost-sensitive: keep Frontier on Haiku 4.5 (good enough), keep Workhorse on gpt-5.4-nano (cheap), and only switch the tier whose output is actually disappointing you.

​

Provider options for each tier

Both tiers support the same four providers:

• Claude (Anthropic). Best output quality at every price point in Runwita’s experience. Fast streaming.
• OpenAI. Strong on the workhorse tier especially. gpt-5.4-nano is the cheapest credible option. gpt-5 and gpt-4.1 work too.
• Ollama (local). Run a model on your own machine, zero API cost, zero data leaves your laptop. Slower and less capable than cloud options. Qwen3-8B is the recommended default if you go this route.
• Custom. Any OpenAI-compatible endpoint. LiteLLM, vLLM, Together, OpenRouter, your own proxy, all work. You set the base URL and model name yourself.

The model dropdown for Claude and OpenAI auto-populates from each provider’s /v1/models endpoint, so you always see what your API key actually has access to. For Ollama and Custom, you type the model name (or pick from /api/tags for Ollama).

​

Privacy implications

You bring your own key. Calls go directly from your machine to the provider you’ve configured, billed against your own account. Runwita doesn’t proxy, store, or see your traffic. Cloud providers (Claude, OpenAI) see the text being processed on each call. That’s the transcript or notes for an extraction, the journey context for a Frontier analysis. They don’t see your full database, just the per-call payload. None of it is used for training (per their respective enterprise terms, governed by your account, not ours). Ollama keeps everything on your machine. Choose Ollama on both tiers if you want zero data leaving your laptop. It’s slower and the output is less polished, but the privacy gain is total. Custom providers (LiteLLM proxy, OpenRouter, etc.) inherit the privacy properties of whatever sits behind your endpoint.

​

What’s next