OverpayingForAIPricing desk

Lesson 8 of 8 · 10 min read · Beginner

Cost control: the cheapest way to get the result you need

Routing, caching, batch, downgrade timing and the red flags of overpaying, with a worked comparison of what one job costs on each OpenAI model.

In this lesson you will

  • Apply model routing so the flagship is the exception, not the default
  • Recognise the five cost drivers behind any OpenAI bill
  • Spot the red flags that mean you are overpaying and know when to downgrade or cancel

Everything in this course leads here. OpenAI's pricing is not hidden, but the defaults (the top model, the $20 plan, the long conversation) are all set to the expensive end. Cost control is mostly the discipline of moving each of those defaults down until the output stops being good enough, then stepping back up one notch.

Model routing: pay for the top model only when the task needs itSend easy work to the cheap tier, hard work to the expensive tier. Most teams find 70–90% of requests are easy.Incoming taskclassify difficultysimple / shortmoderatehard / longSmall modelcheapest tierMid modelworkhorse tierFlagship modelmost expensive tierTypical blended savingAll traffic on the flagship: 100% of cost80% small · 15% mid · 5% flagship~ 20% of costIllustrative only — run your own mix in theAI Cost Calculator with live catalogue prices.
Figure 1.Routing: send each task to the cheapest model that handles it, and escalate only on failure. Most volume lands on mini or nano; the flagship handles the residue.

Rule 1: route by difficulty, not by habit

In ChatGPT, that means using the lighter model in the picker for routine work and saving heavy reasoning for hard problems. In the API it means a literal routing step: classify the request with a nano model, send easy ones to mini, hard ones to standard, and only escalate to GPT-5.5 when a check fails. See how much this changes in the model routing tool.

One job, nine prices: summarise 200 documents (1M input tokens, 100K output tokens)$ per jobGPT-5.58 $5 in + $30 outGPT-5.44 $2.50 in + $15 outGPT-5.4 (Batch)2 half price, asyncGPT-4.12.8 $2 in + $8 outGPT-5.4 mini1.2 $0.75 in + $4.50 outGPT-5 Mini0.45 $0.25 in + $2 outGPT-5.4 Nano0.33 $0.20 in + $1.25 outGPT-5 Nano0.09 $0.05 in + $0.40 outgpt-oss-120b0.05 open-weight
Figure 2.Computed from catalogue prices at the time of writing. The same summarisation job costs $8.00 on GPT-5.5 and $0.09 on GPT-5 Nano. If the nano output is acceptable for even half the documents, routing pays for itself immediately.

Rule 2: know your cost drivers

What actually drives an AI billRoughly ordered by how often they surprise people.1Output length
Output tokens cost 3–5× input. Verbose answers and long reports dominate.
2Re-sent history
Every chat turn re-bills the whole conversation and any pasted files.
3Wrong tier
Flagship model on tasks a small model does identically.
4Retries
Vague prompts → three attempts → three bills.
5Idle subscriptions
Seats and plans nobody used this month.
6Agents & tools
Multi-step loops multiply calls silently.
Figure 3.Five things move an OpenAI bill: model tier, output length, context size (including re-sent history), number of calls (retries, loops) and whether caching or batch applies.
  • Model tier — the biggest lever. A tier down is usually 2–4× cheaper on output.
  • Output length — output costs several times input on every OpenAI text model. Ask for the length you need.
  • Context size — long chats re-send history every turn. New topic, new chat.
  • Call count — retries from vague prompts, loops without limits, polling.
  • Discounts — cached prefixes and Batch jobs; both free money if your workload fits.

Rule 3: downgrade and cancel on evidence

For subscriptions, run the one-week limit test from lesson 4 every quarter. No limit messages on Plus for a month means Go or Free. On Pro, if you did not hit a Plus-level limit most days, drop to Plus and pocket $180 a month. Subscriptions are monthly; there is no penalty for moving down and back up.

For the API, review the usage dashboard monthly by model. If the flagship is more than a small share of your tokens, something is routed wrong. If output tokens dominate, shorten instructions. If the same prefix appears everywhere, restructure for caching.

Finally, prices move. OpenAI has cut per-token prices repeatedly and reshuffled plan entitlements more than once. Anything you decided six months ago deserves a re-check against the pricing tracker and the pricing history. The cheapest plan is the one that matches this month's usage, not last year's.

Knowledge check

You pay for ChatGPT Pro and have not seen a usage limit message in six weeks. What is the cost-correct move?

Lesson FAQ

What is the cheapest OpenAI model?

In our catalogue at the time of writing, gpt-oss-120b and gpt-oss-20b (open-weight) are cheapest, then GPT-5 Nano among the hosted GPT-5 line. Check the live ladder on /models.

How do I cancel ChatGPT Plus?

In the app, open Settings, then Subscription, and choose cancel. Access continues to the end of the billing period. You can resubscribe at any time with no penalty.

Should a small team buy individual Plus plans or Business?

Business costs $20/user annual or $25 monthly with a 2-user minimum, adds admin controls and excludes your data from training. For two or more people doing work with company data, Business is usually the better buy.

Finished reading?

Mark it done to track your progress through the course.

Compare, calculate, decide — for ChatGPT

If our calculators helped you cut down on hidden AI wallet leaks, consider buying us a coffee. A tiny fraction of your savings 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.