Search

Tools

Model Routing & Cascade Calculator

Work out whether a cheap-model-first cascade, a pre-router or advisor mode beats a single model, using your own traffic, accuracy and prices.

Key points from the article

  • A cascade pays off only if a cheap check can tell hard requests from easy ones. Without a trustworthy check, you pay for both attempts and still ship wrong answers.
  • Compare cost per accepted answer, not cost per call. Wrong answers carry rework cost, and a cheap wrong answer is not a saving.
  • The break-even depends on three things: the escalation rate, the cost of the check, and the cost of a missed escalation.
  • Confidence signals differ. Deterministic validators are the most dependable, self-reported scores the least. Measure your signal on your own traffic.
  • Advisor mode keeps a cheap executor in charge and consults a stronger model for advice. Whether it saves money depends on success rates and how long the transcript grows.

Read the full argument and the worked examples inModel Routing and Cascades: Cheap Model First, Escalate on Confidence.

Work through the calculator

Five steps. Each step asks for one thing. Results update as you type.

Step 1 of 5: Start

Step 1 · Start

This tool answers one question: is a cheaper model, or a cheap-then-escalate cascade, cheaper per correct answer than a single strong model? You will enter your workload, prices and accuracy, then read the verdict.

Presets set the routing inputs in steps 2 and 3. Advisor inputs stay as you set them.

The steps ahead:

  1. Workload and prices
  2. Difficulty mix, accuracy and the confidence check
  3. Results: cheapest strategy per accepted answer, and break-even
  4. Advisor mode, for multi-step tasks only
Method, reference points and sources

Published reference points (Anthropic)

These are the figures Anthropic reports for advisor mode. This tool does not reproduce them. They depend on token counts, prompts and effort levels that are not published here. Use them only as a sense of scale.

  • SWE-bench Multilingual, Sonnet executor with Opus advisor: +2.7 percentage points over Sonnet alone, and 11.9% lower cost per agentic task. Averaged over five trials of 300 problems.
  • BrowseComp, Haiku executor with Opus advisor: 41.2% versus 19.7% for Haiku alone (1,266 problems, one attempt each). The blog says this setup costs 85% less per task than Sonnet alone.
  • Docs, hard reasoning benchmark (n = 40 per setting): at max_tokens 2048, mean advisor output is about 630 to 840 tokens with about 0% truncated. At 1024, it is about 370 to 480 tokens with about 10% truncated.

Sources: The advisor strategy (Claude blog);Advisor tool (API docs).

How the routing math works

Each request is classed as easy, ambiguous or hard, with a weight from the mix. A request starts at tier 1. Its answer is correct with probability p from the accuracy table. The check escalates with probabilityp × (right answers escalated) + (1 −p) × (wrong answers caught). Escalations go to the next tier, which repeats the same logic. The last tier is always accepted.

With a perfect check and the article's example prices (1, 5 and 20 units), this reproduces the article's worked numbers: 4 units for the support triage mix, 3.9 when hard requests skip tier 1, 9 for the 30/40/30 mix, and 19 once a judge costing 10 runs on every request. The baseline is 20.

Cost per accepted answer = (cost per request + (1 − quality) × rework cost) ÷ quality. It treats a cheap wrong answer as a cost, not a saving.

How the advisor math works

The executor runs every turn at its price. Each turn's input is the starting context plus the tokens added so far. Advisor calls are spaced evenly through the task. Each call is billed on the full transcript at the advisor price, with optional caching. The first call writes the prefix at the write multiplier. Later calls read the previous prefix at the read multiplier and write only the new tokens. Advisor output is capped at max_tokens (minimum 1024). The tool does not model advice tokens being added back to the executor's context, so it slightly underestimates long tasks.

Disclaimer: Prices and multipliers are illustrative placeholders, not current list prices. Accuracy and quality values are not measured results. Check the latest rates and run your own evaluation before you make a decision.