The Default Mistake
When you first deploy AI agents, the natural instinct is to pick one model and use it for everything. It is the path of least resistance. One API key, one provider, one configuration. It works. Until the bill arrives.
I learned this the hard way. On the first real day of running my Realm, one of my Characters burned through $7 in API credits in a single afternoon. Not because it did anything wrong — it was doing exactly what I asked. It was running a bulk data export through Claude Sonnet. Five million tokens of output at $15 per million. The math was simple. The bill was not.
The problem was not the model. The problem was that I had assigned a premium creative-writing model to a task that needed speed and volume, not elegance. It was like hiring a novelist to fill out spreadsheets.
The right model is not the most powerful model. It is the model whose strengths match the work the Character actually does.
How Characters Think Differently
In REALM, every Character belongs to one of five Classes. Each Class defines what the Character naturally does. And what they naturally do determines what kind of model they need.
A Bard writes content — blog posts, product descriptions, brand voice, social media. The output is public-facing. Quality is visible. A cheap model that writes bland, generic copy damages your brand. This Character needs the best writing model you can afford.
A Warrior executes — listing products, processing data, moving cards on a board, formatting templates. The output is functional, not creative. Speed and reliability matter more than prose. Running this Character on a premium writing model is wasting money on capabilities the work does not need.
A Mage analyzes — financial models, strategic tradeoffs, risk assessment. The output is analytical. It needs reasoning depth, not creative flair. A fast model with strong reasoning handles this well at a fraction of the cost.
A Hunter researches — competitors, market trends, opportunities. They move fast, scan broadly, and report back. Speed matters. Access to real-time data matters. The model that can search the web natively and return structured findings quickly is better than one that writes beautifully but slowly.
A Cleric coordinates — managing the daily brief, tracking the Quest Board, keeping the Player informed. This Character talks to the Player every day. If the daily experience feels flat, the whole system feels flat. Quality of conversation matters here — but not the same kind of quality as creative writing.
My Current Setup
I run Niflheim Records — a physical collector store for vinyl, collectibles, and merch — with 9 AI Characters deployed across seven Zones. Here is how I currently assign models, and why.
| Character | Class | Provider | Why |
|---|---|---|---|
| Freyja | Bard | Claude | Brand voice and content creation. Writing quality is non-negotiable. |
| Fricka | Cleric | Claude | Daily coordination. The Scroll, the Quest Board, priorities. My most frequent conversation. |
| Mime | Warrior | Claude | Code and integrations. Precision matters for technical work. |
| Wotan | Mage | Grok | Strategic analysis. Fast reasoning with a massive context window. |
| Erda | Mage | Grok | Financial analysis. Structured, pattern-heavy work. Speed over style. |
| Loge | Hunter | Grok | Competitor research. Built-in web and X search. Fast scouting. |
| Brünnhilde | Warrior | Grok | Execution and bulk operations. Speed and volume. Never run bulk work on a premium model. |
| Isolde | Cleric | ChatGPT | Personal scheduling and life balance. Warm conversations. Burns existing credits. |
| Saga | Mage | Qwen 3.5 (local) | REALM auditing. Reads the entire Codex. Zero cost for heavy token usage. |
Four providers. Three Characters on Claude for quality-critical work. Four on Grok for speed and analytical work. One on ChatGPT to use existing credits. One running locally at zero cost for token-heavy auditing.
What This Costs
The total estimated cost for running nine agents this way is somewhere between $20 and $30 per month. That is less than a streaming subscription. For a full team of specialists working alongside me every day.
If I had run all nine Characters on Claude Sonnet, the same workload would cost roughly $150 to $200 per month. The difference is not marginal — it is the difference between a sustainable system and one that bleeds money every time an agent does something useful.
Same 9 agents. Same work. Multi-provider: ~$25/month. Single premium provider: ~$175/month. The model assignment is the cost lever, not the number of agents.
The Principles Behind the Assignment
After running this system for real, three principles emerged.
Match the model to the Class, not the Character. A Warrior needs speed regardless of their personality. A Bard needs writing quality regardless of their Zone. The Class tells you what kind of work the Character does. The kind of work tells you what kind of model it needs.
Never run bulk operations on a premium model. This is the single rule that prevents cost surprises. If a Character needs to process thousands of records, export data, or do any kind of high-volume work — that Character should be on the cheapest model that can do the job reliably. Volume times cost-per-token is the equation that breaks budgets.
Invest in quality where the Player feels it. Your daily Cleric, your content Bard, your code specialist — these are the Characters whose output directly impacts your experience and your customers. Saving money on them saves pennies and costs quality. Invest where it matters, economize where it does not.
The Platform That Makes This Possible
This multi-provider setup runs on OpenClaw — an open-source agent platform that connects AI models to messaging tools like Slack, WhatsApp, and Telegram. Each Character runs as an independent agent with their own model, their own workspace, and their SOUL file as the system prompt.
OpenClaw supports multiple providers natively — Anthropic, OpenAI, xAI, Google, and local models through Ollama. You assign a different model to each agent in one configuration file. When a provider has issues, OpenClaw falls back to the next one automatically. The Realm stays online even if one provider goes down.
This is not an endorsement. It is what I use. REALM is platform-agnostic — the framework works on any infrastructure that can host agents with system prompts. But the ability to mix providers per Character is what makes the cost optimization real. If your platform locks you into one provider, you lose this lever.
What I Would Tell Someone Starting Today
Start with one model for everything. Get the system working. Understand what each Character actually does in practice, not in theory. Then optimize.
Your first week will teach you which Characters generate the most tokens. Your first month will teach you which Characters need quality and which need speed. Your first bill will teach you where the money goes. All of that information is what you need to make intelligent model assignments.
Do not optimize before you have data. But do not ignore the data once you have it.
The company of one cannot afford to waste money on models that do not match the work. But it also cannot afford to run everything on the cheapest option and wonder why the output feels flat. The answer is in the middle — the right model for the right Character, calibrated by experience.