Choosing Models, Effort & Trade-offs

Providers offer tiers: frontier models (best reasoning, highest cost), balanced models, and small/fast models (classification, extraction, routing). Many current models also expose a reasoning effort control — higher effort thinks longer (better on hard problems, more tokens and latency). Start with the most capable model to establish a quality baseline, then measure cheaper models or lower effort on your eval set before downgrading routes. Keep model IDs in configuration, not scattered in code, so upgrades are one change plus an eval run.

Model tiers

MODELS = {                                   # one place to upgrade models
    "agent": {"model": "claude-opus-5-5", "effort": "high"},
    "summarize": {"model": "claude-opus-5-5", "effort": "low"},
    "classify": {"model": "claude-haiku-5-5", "effort": "low"},
}

def call(route: str, **kwargs):
    cfg = MODELS[route]
    return client.messages.create(model=cfg["model"], output_config={"effort": cfg["effort"]}, **kwargs)

const MODELS = {
  agent: { model: "claude-opus-5-5", effort: "high" },
  classify: { model: "claude-haiku-5-5", effort: "low" },
} as const;

const r = await client.messages.create({
  model: MODELS.classify.model, max_tokens: 256,
  output_config: { effort: MODELS.classify.effort },
  messages: [{ role: "user", content: `Classify sentiment: ${review}` }],
});

Common pitfalls

• Picking the cheapest model before measuring quality.
• Hard-coded model strings across the codebase.
• Comparing cost per request instead of cost per successfully completed task.

Key points

• Baseline with the best model, then measure cheaper options
• Route by task
• Model IDs in config

Further reading: Claude models overview