MiroFish API Costs: How to Cut Spend by 90%

How MiroFish Burns Through Tokens

Each agent in each round makes at least one LLM call to think, react, and update its memory. Graph construction, persona generation, environment setup, and report generation all use additional calls. A 40-round simulation with 100 agents generates roughly 4,000-5,000 LLM calls.

Cost per Model

GPT-4o runs about $0.50-2.00 per full simulation. GPT-4o-mini drops that to $0.01-0.05. DeepSeek V3 is even cheaper at roughly $0.005-0.02. Google Gemini Flash has a generous free tier that can cover lightweight simulations at zero cost.

The Hybrid Model Strategy

Use a strong model (GPT-4o-mini) for quality-critical tasks like persona generation and report writing, and a cheap model (Gemini Flash or DeepSeek) for high-volume tasks like individual agent interactions. MiroFish supports a secondary LLM configuration in .env for exactly this purpose.

Limit Simulation Rounds

Start with fewer than 40 rounds as the MiroFish team recommends. Early runs should be about calibration — checking whether your seed material is clean and your agents make sense. You don't need 100 rounds to get useful predictions.

Run Locally with Ollama

The MiroFish-Offline fork by nikmcfly lets you run entirely locally with Ollama and Neo4j. Zero API cost, but you need decent hardware — at least 16GB RAM and a GPU helps. Agent quality depends heavily on model size.