Dirt cheap

The brief: spend as close to nothing as possible and still get real work done. Point every role at the cheapest capable models, turn thinking off where it doesn't pay, cap output, and put a hard --budget under it all. Great for high-volume chores — lint sweeps, docstrings, mechanical migrations, triage-the-backlog — and for getting a feel for Stella before spending real money.

The config

{
  "agent_engine_config": {
    "pipeline_worker_model": "deepseek/deepseek-chat",
    "pipeline_judge_model": "zai/glm-5.2",
    "pipeline_triage_model": "deepseek/deepseek-chat",
    "default_model": "deepseek/deepseek-chat",

    "allowed_models": ["deepseek/deepseek-chat", "zai/glm-5.2"],

    "auto_mode": "off",
    "effort_auto": "off",
    "reasoning_auto": "off",

    "agents": {
      "worker": {
        "effort": "low",
        "reasoning": "off",
        "params": { "max_tokens": 8192 }
      },
      "judge": { "effort": "medium", "reasoning": "on" },
      "triage": { "reasoning": "off" }
    }
  }
}

Why these choices:

• DeepSeek as worker — $0.27/$1.10 per Mtok with a $0.07 cached rate: the catalog's price floor among hosted models. Thinking off and max_tokens capped: mechanical work doesn't need deliberation, and a runaway answer can't run far.
• GLM-5.2 as judge — still cheap, still a different family than the worker, with thinking on: the judge is the one place this profile keeps its brain switched fully on. The whole judge call costs a fraction of a cent.
• No autos — effort_auto would raise the worker to medium; this profile wants the floor.

Run it with a hard cap

# A quarter, and not a cent more.
stella --budget 0.25 run "add docstrings to every public function in src/api/"

# Chores at fleet scale — the budget divides across workers and caps the whole run.
stella --budget 1.00 fleet \
  "fix clippy warnings in stella-store" \
  "fix clippy warnings in stella-tools" \
  "normalize log messages to sentence case"

Always pass --test-command when there are tests — the deterministic fail→pass flip is free verification that skips the judge call entirely on the happy path.

What it costs

The $0.00 tier: local models

Actually zero: serve a model yourself (Ollama, vLLM, LM Studio, llama.cpp) and pay in electricity —

stella --model local/llama3.3 --base-url http://localhost:11434/v1 \
  run "explain what main.rs does"

Quality depends entirely on what you serve; see Local servers. A pragmatic split: local model as worker, one cheap hosted model as judge, so verification keeps an independent set of eyes.