Which AI Model for Which Agent? How to Choose Smartly

The Problem With "Best Model for Everything"

When we built our first multi-agent setup, the configuration was simple: all agents on Claude Opus. The strongest, the most expensive — better safe than sorry.

Then the API invoice came after week one. For six agents running around the clock, the number was... significantly higher than expected.

The realization: Not every task needs the most powerful model. When Alex checks my calendar and tells me whether I have a meeting tomorrow — that doesn't need 200-token-per-second intelligence. But Peter, our coding agent, reviewing complex TypeScript architectures — he genuinely needs the best available.

The result of our overhaul: 60% fewer API costs with equal or better quality.

---

The Core Idea: Classify Your Tasks

Before assigning models, you need to understand what each agent actually does. We split our agents into three categories:

Category 1: Reasoning-Intensive Tasks

These tasks require deep thinking, multi-step inference, code analysis, or creative quality work.

Examples:

Recommended models: Claude Opus 4.5+, GPT-4o, Gemini Ultra

Category 2: Structured, Rule-Based Tasks

These tasks follow clear patterns. Input is structured, output is predictable, error rates are low.

Examples:

Recommended models: Claude Sonnet 4.5, GPT-4o Mini, Gemini Flash

Category 3: Simple Execution

Tasks where the model mostly acts as an interface: receive command, call tool, return result.

Examples:

Recommended models: Claude Haiku, GPT-4o Mini, Gemini Flash 8B

---

How to Configure Models in OpenClaw

OpenClaw lets you set the model per agent in `~/.openclaw/openclaw.json` (or `openclaw.json5`). The configuration looks like this:

```json

{

"agents": {

"sam": {

"model": "anthropic/claude-opus-4-5",

"workspace": "/home/sam/.openclaw/workspace"

},

"peter": {

"model": "anthropic/claude-opus-4-5",

"workspace": "/home/peter/.openclaw/workspace"

},

"maya": {

"model": "anthropic/claude-sonnet-4-5",

"workspace": "/home/maya/.openclaw/workspace"

},

"alex": {

"model": "anthropic/claude-haiku-3-5",

"workspace": "/home/alex/.openclaw/workspace"

},

"iris": {

"model": "anthropic/claude-sonnet-4-5",

"workspace": "/home/iris/.openclaw/workspace"

},

"atlas": {

"model": "anthropic/claude-opus-4-5",

"workspace": "/home/atlas/.openclaw/workspace"

}

}

}

```

Alternatively, you can set the model per agent via an environment variable in the Docker Compose file:

```yaml

services:

alex:

image: openclaw/agent:latest

environment:

- OPENCLAW_MODEL=anthropic/claude-haiku-3-5

- OPENCLAW_AGENT_NAME=alex

volumes:

- /home/alex/.openclaw/workspace:/workspace

```

Both methods work. We prefer the JSON configuration because it documents all agents centrally.

---

Our Real Setup: The 6-Agent Model Matrix

Here's exactly what we run — no theory, the actual setup:

| Agent | Role | Model | Reason |

|-------|------|-------|--------|

| Sam (Team Lead) | Delegation, planning, blog | Claude Opus 4.5 | Complex coordination, creative writing |

| Peter (Coding) | PR reviews, tests, bugs | Claude Opus 4.5 | Architecture understanding, reasoning |

| Maya (Marketing) | Copy, campaigns, SEO | Claude Sonnet 4.5 | Good quality, 3× cheaper than Opus |

| Alex (Admin) | Calendar, reminders, tasks | Claude Haiku 3.5 | Structured tasks, no depth needed |

| Iris (Research) | Research, synthesis, reports | Claude Sonnet 4.5 | Reasoning + cost balance |

| Atlas (CEO Support) | Strategy, reports, letters | Claude Opus 4.5 | CEO output must be flawless |

Cost comparison (estimated, moderate usage):

---

How to Tell When an Agent Has the Wrong Model

Signs of "model too weak":

Signs of "model over-provisioned":

The fix: take a short look each week at the ratio of token consumption to output quality. With Alex, we noticed after two weeks that Haiku handles 95% of his tasks just fine.

---

Dynamic Model Switching: The Next Level

Advanced setups can switch models within an agent based on context. OpenClaw supports this via the session status override:

```

# In chat with the agent:

/model anthropic/claude-opus-4-5

# Or programmatically in a cron job instruction:

"For this task use /model anthropic/claude-opus-4-5 — complex analysis needed."

```

We use this rarely — mainly when Iris gets a particularly complex research assignment and needs to briefly switch to Opus. The default config stays Sonnet.

---

Mixing Providers: OpenAI + Anthropic + Gemini

OpenClaw supports multiple providers simultaneously. That means: you don't have to commit to one vendor.

```json

{

"agents": {

"maya": {

"model": "openai/gpt-4o-mini"

},

"alex": {

"model": "google/gemini-flash-1.5"

}

}

}

```

Important: each provider needs its own API key in `.env`:

```bash

ANTHROPIC_API_KEY=sk-ant-...

OPENAI_API_KEY=sk-...

GOOGLE_API_KEY=...

```

We experimented with mixed providers but ultimately stayed with Anthropic for all agents — consistent quality and a single invoice is operationally simpler.

---

Practical Guide: The Three-Question Model Decision

If you're unsure which model is right for an agent, ask these three questions:

1. Does the agent need to weigh contradictory information?

→ Yes → At least Sonnet, preferably Opus

2. Are the agent's tasks mostly predictable and structured?

→ Yes → Haiku or Flash is enough

3. Does a human see the output directly (CEO, customer, public)?

→ Yes → No compromise: Opus

These three questions helped us move our setup from "all Opus" to a thoughtful mix.

---

The Complete Setup

The exact configuration — openclaw.json, Docker Compose with model variables, and the criteria by which we selected each model — is documented in the OpenClaw Setup Playbook.

Including the monitoring setup we use to track token consumption per agent and detect when a model is over- or under-performing.

18 chapters, based on real production experience.

Fully available in German too. 🇩🇪