How do I route AI models by cost?

Set a cheap default (Grok 4.5, Gemini 3.6 Flash, or DeepSeek V4 Flash) for the ~80% of easy, verifiable steps, then escalate to a flagship (Claude Opus 5 or GPT-5.6 Sol) on retries, low-confidence outputs, or hard tasks. Verify cheap-tier output programmatically so misses are caught cheaply.

How much can model routing save?

A lot. If 80% of steps run on a cheap model at ~$0.08/task and 20% escalate to a flagship at ~$0.28/task, your blended cost is ~$0.12/task versus $0.28 running everything on the flagship — roughly 55% savings with minimal accuracy loss on hard steps.

When should a step escalate to a flagship model?

Escalate on: a failed verification (test/lint/schema), low model confidence, a retry after error, or a task flagged as hard (multi-repo refactor, ambiguous spec). Keep everything else on the cheap default.

Quick Answer

How to Route AI Models by Cost in 2026 (Setup Guide)

Published: July 31, 2026

The Short Answer

Model routing means defaulting to a cheap model and escalating to a flagship only when needed. Done right, it cuts costs ~50%+ with almost no accuracy loss. The trick is a good escalation trigger and programmatic verification so cheap-tier misses are caught cheaply.

The Three-Tier Setup (2026)

Tier	Model	Price ($/MTok)	Role
Cheap default	Gemini 3.6 Flash / Grok 4.5	$1.50/$7.50 · $2/$6	~80% of steps
Value quality	GPT-5.6 Luna / Gemini 3.1 Pro	$1/$6 · $2/$12	mid-tier fallback
Flagship	Claude Opus 5 / GPT-5.6 Sol	$5/$25 · $5/$30	hard/low-confidence

(In the EU, drop Grok 4.5 — it’s not available there.)

Step-by-Step

Pick a cheap default. Gemini 3.6 Flash (65% fewer output tokens) or Grok 4.5 (~2x token efficiency). For open-weight/self-host, DeepSeek V4 Flash ($0.14/$0.28).
Add programmatic verification. After each cheap-tier output, run a check: unit tests, linter, JSON-schema validation, or a rules engine. This is what makes cheap routing safe.
Define escalation triggers. Escalate to a flagship when: verification fails, the model reports low confidence, a retry follows an error, or the task is pre-tagged hard (multi-repo refactor, ambiguous spec).
Cap retries. One cheap retry, then escalate — don’t burn tokens looping on the cheap tier.
Log and tune. Track escalation rate. If >30% of steps escalate, your default is too weak or your tasks are harder than assumed.

The Savings Math

Say 80% of steps stay on a cheap model (~~$0.08/task) and 20% escalate to Opus 5 (~~$0.28/task):

blended = 0.8 × $0.08 + 0.2 × $0.28 = $0.064 + $0.056 = $0.12/task

That’s ~$0.12/task vs $0.28 running everything on the flagship — roughly 55% cheaper with hard tasks still getting frontier accuracy.

Common Mistakes

No verification. Cheap routing without a check just ships more bugs. Verification is non-negotiable.
Escalating too eagerly. If low confidence triggers on everything, you pay flagship prices anyway. Tune the threshold.
Ignoring token efficiency. Gemini 3.6 Flash and Grok 4.5 finish in fewer tokens — count effective cost per task, not sticker rate.

Sources

Artificial Analysis — model comparisons: artificialanalysis.ai
Anthropic — Claude Opus 5: anthropic.com/news/claude-opus-5
OpenAI — GPT-5.6: openai.com/index/gpt-5-6