OpenAI’s most expensive model costs 100 times as much per token as its cheapest. GPT-6 Astra lists at $10 per million tokens in and $50 out. GPT-6 Luna lists at $0.10 in and $0.50 out. Anthropic’s range is narrower but still a tenfold gap, from Claude Haiku 4.5 at $1/$5 to Claude Fable 5.1 at $10/$50. Most small businesses we speak to use one model for everything, which means they are either overpaying for routine work or under-powering the jobs that matter.

Here is how we sort work before we pick a model, and how you can do the same this afternoon.

Three piles: thinking, routine and bulk

Take ten things you asked AI to do last week and tag each one:

• Thinking. Strategy, a delicate reply to an unhappy client, anything with money, contracts or compliance in it. A wrong answer here costs you something real.
• Routine. Drafting an email, summarising a meeting, tidying notes, writing a first pass of a quote. A decent answer is fine and you will read it before it goes anywhere.
• Bulk. Sorting 500 enquiries, tagging invoices, pulling dates and amounts out of documents. Volume is the point and each individual item is low stakes.

Each pile suits a different tier of model.

What the tiers cost, side by side

These are API list prices per million tokens (input/output) as we checked them on 5 October 2026. A token is roughly three-quarters of a word. Prices are in US dollars.

• Top tier: Claude Fable 5.1 at $10/$50 and GPT-6 Astra at $10/$50.
• Mid tier: Claude Opus 5.5 at $4/$20, Claude Sonnet 5.5 at $2/$10, GPT-6 Sol at $2/$10 and Gemini 3.1 Pro at $2/$12.
• Cheap tier: Claude Haiku 4.5 at $1/$5, Gemini 3.1 Flash-Lite at $0.25/$1.50 and GPT-6 Luna at $0.10/$0.50.

Most of the major providers also offer a batch option that takes roughly half off for work that does not need an instant answer, and cached input is cheaper again. Check each provider’s pricing page before you rely on either. Model lines change quickly: Anthropic released Opus 5.5 on 22 September and Sonnet 5.5 on 28 September, and OpenAI released Sol and Luna on 22 September. Google has also announced that its older 2.5 Flash-Lite models shut down on 16 October, so anything still pointed at them needs moving.

The prices come from the vendors’ published rate cards as reported by Lindy’s GPT-6 guide, BenchLM’s Claude pricing tracker and CostGoat’s Gemini pricing guide.

A worked example, not a promise

Say you want an AI to draft replies to 1,000 customer emails a month. Assume each email plus your instructions is about 700 tokens in and each reply is about 200 tokens out. That is 0.7 million tokens in and 0.2 million out in total. At list price:

• GPT-6 Astra or Claude Fable 5.1: about $17
• Claude Opus 5.5: about $6.80
• GPT-6 Sol or Claude Sonnet 5.5: about $3.40
• Claude Haiku 4.5: about $1.70
• GPT-6 Luna: about $0.17

That is an illustration of how the maths works, not a forecast of your bill. Your emails will be longer or shorter, your prompts will differ, and these are tool costs only. The point is the shape: the same job can differ by a factor of 100 in price depending on which model you point at it.

The trade-off: cheaper can mean more retries

A smaller model can miss nuance, need a clearer prompt, or take a second attempt. Every retry costs tokens and your time. A model that is 20 times cheaper but needs three goes to get a usable answer is still cheaper, but one that needs ten goes may not be, and one that quietly gets the answer wrong is worse than either. That is why we do not move a whole workflow to a cheaper tier on the strength of a price list.

How we apply it

1. Default to the mid tier. It handles most drafting and summarising. Move up only when it fails, not before.
2. Reserve the top tier for thinking jobs. We pick one or two a month where the stakes justify it.
3. Send bulk work to the cheap tier, but check ten results by hand before trusting the rest.
4. Run the same ten real jobs on two tiers and compare the answers side by side before changing anything.

If you use a chat app rather than the API, the same idea applies: look for the model picker, set a mid-tier model as your default, and only switch to the top model when a job earns it. Subscription plans are priced differently from the API, so the dollar figures above matter most if you are paying per use or building automations.

Why this matters more as you scale

At one email a day, none of this matters. The gap bites when you automate. A workflow that reads every incoming enquiry, every invoice or every review runs thousands of times a month, and each run carries a price. Choosing the model per job, rather than per business, is the cheapest habit you can build before you connect AI to anything that runs on its own. It also makes your costs easier to predict, because each workflow has one model and one price you can work out in advance on the back of an envelope.

It also protects you when the market moves. Prices and model names have changed several times in the last two months, so a workflow hard-wired to one top-tier model has no easy way to take advantage of a cheaper option, or to move off a model that is being retired. Keep the model name in one place, note why each job uses the tier it does, and review the list whenever a vendor announces a release.

Start here this afternoon

List ten jobs, tag them thinking, routine or bulk, and run one routine job on a cheaper model than you normally use. Compare the two answers. If you cannot tell them apart, you have just found a job you were overpaying for.

We sort every client workflow into these three piles before we choose a model, so the top tier only gets the work that needs it. Which job are you running on a top-tier model that a cheaper one could handle?