ET RU EN

Blog

Ryanair is looking to AI for savings. Does a small business need the most expensive model?

Published

Ryanair expects fuel to cost the company €2 billion more in the next financial year, and among the ways to save money, chief executive Michael O'Leary names wider use of AI in administrative work. But AI comes with a bill too, and its size depends largely on one choice that is easy to make out of habit: which model you use for each task. The short answer for a small business owner in Estonia or a self-employed professional: simple tasks don't need the most expensive model. Drafts, sorting through emails and template texts can go to a cheaper model. The flagship, meaning the strongest and most expensive model, is worth keeping for tasks where a mistake costs more than the difference in price.

Why would an airline turn to AI when fuel gets more expensive?

To save money on the costs the company controls itself. As ERR reports, Ryanair chief executive Michael O'Leary forecasts a €2 billion rise in fuel costs in the next financial year and warns that ticket prices will go up. Among possible cost-saving measures, he names a temporary pay freeze and wider use of AI in administrative work.

Every firm has administrative work, from an airline to a two-person workshop: emails, documents, reports, answers to routine questions. So the same question - where AI can save money - applies to a small firm too.

But AI comes with a bill too. When a model is connected to a firm's work through an API (a programmatic interface used to build AI into services and spreadsheets), payment is usually counted in tokens - small chunks of text: separately for what you send the model and for what it writes back. So saving money on AI starts with the question of which model you assign to each task.

How much does a flagship cost, and is it worth paying less?

It is, when the task is simple. For reference, take Google's latest flagship. According to a post by @_philschmid, Gemini 4 Carbon scores 59 points on the Artificial Analysis v5 index and costs the same as the previous model, Argon: $2 for input tokens and $10 for output tokens. The post doesn't specify the unit, and the price is marked with an asterisk. For now only trusted partners are testing the model; developers have been promised access as soon as possible.

At the other end are the cheaper models. In the estimate of an analyst who posts on X as @bindureddy, DeepSeek is 20 times cheaper than Opus and already handles 70% of tasks. The author's conclusion: companies overpay heavily when they give simple work to expensive models. The figures come without a source, so this is an opinion, and it is worth testing on your own tasks.

The thinking behind it is clear enough, though. A payment reminder, a draft post, sorting incoming mail by topic - work like this usually doesn't need the most capable model available. So before each task it is worth asking yourself: does it need the best result on the market, or will a good one do?

Then why not move all the work to a cheap model?

Because every model has its own strengths, and the hardest tasks are still handled better by closed flagships. Artificial Analysis has switched to private benchmarks (tests whose tasks are not published), and the new measurements showed that open models - those whose weights are publicly released and which you can run yourself - lag further behind closed ones than earlier results suggested. As Ethan Mollick writes, the result should be treated with caution, but it matches the general impression: there is no open model at Fable's level yet.

The cheap end of the market is moving too. As Perplexity's chief executive Aravind Srinivas writes, a company called Particle is launching: it tunes open models so that they use fewer tokens and cost less to run. Perplexity worked with the team before the launch and plans to collaborate with it on search and on routing tasks between models.

So open models are getting cheaper, but closed ones still lead on complex tasks. And even among closed models, each has its own strengths. Here is the latest LiveBench ranking, as @bindureddy reports:

  • Fable 5.1 and Astra are at the top.
  • Opus 5.5 is close behind and is the best at design.
  • Argon beats Terra and is strong in data analysis.
  • A Google model has overtaken open models for the first time; the best of the open models is that same DeepSeek.

What should go to a cheaper model, and what is worth paying for?

A working rule: the more a mistake costs and the less you check the result before it goes out, the stronger the model the task needs. Tasks you will read and correct yourself anyway can go to a cheaper model. Where a mistake hits your money or your clients and checking everything by hand takes a long time, you need a model that is strong at that particular task. Here is how it looks for typical small business tasks:

  • A draft post, email or service description - a cheaper model: the text will go through your edits anyway.
  • Processing incoming email (sorting by topic, summarising a long thread, listing a client's questions) - a cheaper model: a miss here is obvious straight away.
  • Template texts (a payment reminder, an appointment confirmation, an answer to a common question) - a cheaper model: you check the template once, and it writes from it after that.
  • Analysing the figures you use to decide where money goes - a model strong in data analysis: on LiveBench, Argon stands out here.
  • Design, layouts, visual presentation - a model strong in design: in the same ranking, Opus 5.5 is the best here.
  • A long task with many steps, where the model moves from step to step on its own - a flagship: as @OfficialLoganK writes, Google calls Gemini 4 Carbon the best specifically at long multi-step tasks.

Where will this rule let you down, and what should you do then?

The rule has three weak spots, and each has a workaround.

First, the figures on cheap models are still only an estimate. ‘20 times cheaper’ and ‘70% of tasks’ are one author's opinion without a source. What to do: test it on your own material. Give the same task to a cheaper model and to a flagship, and compare both answers with what you would have done yourself.

Second, the picture changes quickly. Gemini 4 Carbon came out just two weeks after Argon, and the change of methodology at Artificial Analysis noticeably shifted the scores of open models. As Nathan Lambert writes, the industry needs narrower indices because models are used for very different tasks. What to do: choose a model for the specific task and re-check your choice when a new version comes out.

Third, the list price and the actual bill can differ. The post on Gemini 4 Carbon gives no unit for the price but does carry an asterisk, and for now only trusted partners can use the model. What to do: compare models using the official pricing of the service you actually work through, and look at the total bill for the month.

To sum up: simple work is fine with a cheaper model, complex work with costly consequences needs a flagship, and money leaks away where one model is used for everything. One step for this week: list the tasks you already give to AI and, next to each, note who checks the result before it goes out. Tasks you check yourself are the first candidates for a cheaper model. If the question goes beyond choosing a model and is about how to restructure the way your business works, have a word with MOST's experts.

Sources

Questions and answers

Which AI model should I choose for simple work tasks?
For drafts, sorting emails and template texts, a cheaper model is enough. In the estimate of analyst @bindureddy, DeepSeek is 20 times cheaper than Opus and already handles 70% of tasks. The figures come without a source, so test it on your own task: give it to two models and compare the results.
How much does a flagship AI model cost?
As a reference point: Google's Gemini 4 Carbon has a stated price of $2 for input tokens and $10 for output tokens. The post doesn't specify the unit, the price is marked with an asterisk, and for now only trusted partners are using the model. It is best to check the exact terms in the official pricing once access opens.
Have open models caught up with closed ones yet?
Not yet. After Artificial Analysis switched to private benchmarks, the gap turned out to be bigger than earlier measurements showed. Ethan Mollick advises treating the result with caution, but also writes that it matches the general impression: there is no open model at Fable's level yet.
How can I tell that a task needs an expensive model?
Look at the cost of a mistake and at who checks the result. If you read and edit the text before sending it anyway, a cheaper model will do. If a mistake hits your money or your clients and checking everything by hand takes a long time, choose a model that is strong at that particular task: on LiveBench, for example, Opus 5.5 is the best at design and Argon is strong in data analysis.
Why are companies looking to AI for savings now?
Because other costs are rising. Ryanair, for example, expects fuel to cost €2 billion more in the next financial year, and among its cost-saving measures it names wider use of AI in administrative work, ERR reports.
Will cheaper models get even cheaper?
Work on that is already under way. A new company, Particle, tunes open models so that they use fewer tokens and cost less to run, and Perplexity plans to work with it. So it is worth reviewing your choice of model from time to time.

← All blog posts