Balanced

The brief: most of the quality, a fraction of the bill. The insight this profile is built on: the worker emits ~85% of your tokens, the judge emits almost none — so spend down on the worker and keep a flagship judge as cheap insurance. This is the sensible default for day-to-day work.

The config

{
  "agent_engine_config": {
    "pipeline_worker_model": "zai/glm-5.2",
    "pipeline_judge_model": "anthropic/claude-fable-5",
    "pipeline_triage_model": "deepseek/deepseek-chat",
    "default_model": "zai/glm-5.2",

    "allowed_models": [
      "zai/glm-5.2",
      "anthropic/claude-fable-5",
      "openai/gpt-5.5",
      "deepseek/deepseek-chat"
    ],

    "auto_mode": "off",
    "effort_auto": "on",
    "reasoning_auto": "on"
  }
}

Why these choices:

• GLM-5.2 as worker — the price/performance sweet spot of the catalog: excellent coding at $2.20/Mtok out vs. a flagship's $15. On a Z.ai coding-plan subscription (one settings line) the marginal cost drops further still.
• Fable 5 as judge. A flagship verdict costs cents per verification and is naturally cross-family to the GLM worker. The judge is where you don't save. (Prefer OpenAI? openai/gpt-5.5 judges equally well here.)
• effort_auto / reasoning_auto on. The autos encode the sensible mapping — judge high, worker medium, triage low; thinking on everywhere except triage — and you stop thinking about it. That's their contract.
• Swap-in: worker gemini/gemini-3-pro when your tasks need the 1M-token window (whole-repo comprehension) at the same output price class as GPT-5.5.

Feeling indecisive about the judge? Turn auto_mode: "on" and let stella pick the best cross-family judge from allowed_models per run.

Run it

stella run "add pagination to the /orders endpoint and update its tests" \
  --test-command "pnpm test orders"

stella --budget 3.00 goal "clippy is clean and the test suite passes"

What it costs

Roughly 5–7× cheaper than the maximum-quality profile on the same work, with the same class of judge watching the result. If runs start needing extra revision rounds on your hardest tasks, promote just the worker to Fable 5 for those runs:

stella --model anthropic/claude-fable-5 run "the hard one"

— an explicit --model outranks the settings for that invocation, and everything else (judge, triage, autos) stays put.