OverpayingForAIPricing desk

Lesson 4 of 6 · 13 min read · Beginner

Pick the cheapest model that still passes

Run one prompt across four price tiers, score the answers yourself, and turn the result into a routing rule. This is the lesson that saves you the most money.

In this lesson you will

  • Run identical prompts across several models
  • Score on acceptance, not on which answer sounds nicest
  • Write a routing rule you can defend

Almost everyone picks a model by reputation and then never revisits it. The catalogue spans $0.03 to $2.00 per 1M input tokens — a 67× range. If a model at the bottom of that range passes your quality bar, reputation just cost you 98% of your bill.

The only way to know is to test on your own work. Here is the whole method in one command.

The same prompt across four price tiers

One real task — rewriting a support reply — run across the price ladder.

recorded session — not a live shell0 / 2
  1. for m in qwen/qwen3.7-flash openai/gpt-oss-120b google/gemini-2.5-flash-lite anthropic/claude-sonnet-5; do echo "--- $m" curl -s https://openrouter.ai/api/v1/chat/completions \ -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" \ -d "$(jq -n --arg model "$m" --rawfile p prompt.txt '{model:$model, max_tokens:200, messages:[{role:"user",content:$p}]}')" \ | jq -r '.choices[0].message.content' doneFour answers, four price points. Read them before you look at the prices — otherwise you will find quality wherever you expect it.
  2. # what each one would cost for 2,000 replies a monthAssuming ~180 input and ~120 output tokens per reply. The frontier model is roughly 74x the cheapest — and every one of those four replies would have been sent.

Your own numbers will differ — token counts depend on your text, and rates move. The arithmetic is what transfers.

Score on acceptance, not on taste

The Sonnet answer is the best writing. That is not the question. The question is: would you have sent it? If the answer is yes for all four, you are looking at a 74× price difference for a preference. If the cheap one would have needed an edit, then it is not cheap — add your own minutes at your own rate and re-do the sum.

Score your own runs like this. One row per model, ten real tasks each.
Model$ / 1M inSent as-isNeeded an editUnusableReal cost per accepted reply
qwen/qwen3.7-flash$0.03730$0.00002 + 3 edits
openai/gpt-oss-120b$0.037820$0.00003 + 2 edits
google/gemini-2.5-flash-lite$0.10910$0.00007 + 1 edit
anthropic/claude-sonnet-5$2.001000$0.00156 + 0 edits

Knowledge check

A cheap model passes 7 of 10 tasks; a frontier model at 60× the price passes 10 of 10. Editing a failure costs you 4 minutes. When is the cheap model still the right choice?

Lesson FAQ

How many tasks should I test on?

Ten real ones beats a hundred synthetic ones. You are looking for a pass rate you would bet on, and ten is enough to see an obvious difference. If it is close at ten, go to thirty.

Do cheaper models get worse over time?

The named model does not change, but what counts as cheap does — new releases keep landing below the old floor. Re-run this test roughly quarterly; it takes twenty minutes and the price ladder moves a lot.

Finished reading?

Mark it done to track your progress through the course.

Save progress across devices

Get a private link that restores your lessons on any device. Email is optional and only used to send you the link.

Compare, calculate, decide

If our calculators helped you cut down on hidden AI wallet leaks, thanks for using them. A tiny fraction of your savings is what keeps our pricing indexes updated daily.

Not sure which AI is cheapest for your use case? Find out in 30 seconds — no signup required.

AI cost intelligence

Stop overpaying for AI tools

Join the OverpayingForAI list for pricing updates, cheaper alternatives, and practical buying guidance.

Now tracking 50+ AI tools, models, platforms, subscriptions, coding tools, and automation products.

We use your email only for OverpayingForAI updates. Unsubscribe anytime.