--- title: How to choose an AI model for each kind of work description: Eight practices for an AI bill that follows the work. Sort the work into kinds, start a tier down, test on your own tasks, and read the bill by who spent it. url: https://salexhub.com/blog/choose-an-ai-model-for-each-kind-of-work category: Guides author: SalexHub team published: 2026-11-13 last_updated: 2026-11-13 --- # How to choose an AI model for each kind of work Eight practices for an AI bill that follows the work. Sort the work into kinds, start a tier down, test on your own tasks, and read the bill by who spent it. Most businesses choose their AI once. Everyone gets the same assistant, the same model answers everything, and nobody looks at the choice again. That habit has a price. On the makers' own price lists, the strongest model costs up to a hundred times the lightest, and the makers' own guides say to keep the lightest setting that does the job. We read 38 published sources on what this costs and what choosing gets back. The full review is in [Cheaper only if you choose](https://salexhub.com/research/cheaper-only-if-you-choose). The gain from choosing is real, and it's smaller than the headlines. Most of it comes from knowing what kind of work you're handing over. Here are eight practices that follow from that. ## 1. Sort the work before you look at models The best-supported finding in the research is that the kind of work predicts which model is worth paying for. So start with a list of the jobs you give AI every week, and put each one in one of four groups: - **Sorting and reading out.** Tagging requests, pulling an amount or a date out of an email, checking a form. - **Writing from what you supply.** Drafts, summaries, replies based on a document or a record you've given it. - **Working in steps.** Calculations, plans, anything where one step depends on the last. - **Long tasks nobody watches.** Work that runs for many steps and reports back at the end. The first two groups are where a lighter model does the job. ## 2. Start one tier down The makers don't tell you to start at the top. OpenAI's guide says to "keep the lightest setting that meets your quality bar". Google's says to "pick the lowest level that meets your quality bar". Anthropic starts its readers on its second tier and names its lightest model for simple tasks. Give the first group to the lightest model your supplier offers and the second to a middle one. Move a job up a tier when you've seen it fail, and only that job. ## 3. Test on twenty of your own tasks Every maker's guide ends the same way: its table is a starting point, and your own work decides. You don't need a lab for this. Take twenty real tasks of one kind, run each on two models, and compare the results side by side. Count the time you spent fixing the weaker answers. Keep the lighter model if it passes. Twenty tasks of your own will tell you more than any benchmark. The published savings, which run from nothing to 98%, were all measured on test sets. Nobody has published what a business saved. ## 4. Price the finished task A lighter model isn't cheaper when it fails. On 175 simulated office tasks of many steps, one maker's lighter model cost USD 0.6 a task and finished 11.4% of them. Its stronger model cost USD 4.2 and finished 30.3%. A failed task still gets paid for, and then someone does it again. So the fourth group, long tasks nobody watches, gets the strongest model you have. That's where the money for it is well spent. ## 5. Turn reasoning up for steps and down for recall Many models can think before they answer, and that thinking is billed as if the model had written it, whether or not you see it. In one study of hard test questions, a median 94.7% of the cost of what the models generated went on thinking. It pays on work in steps: calculation, code, planning. It pays little on questions of recall or on a short edit, and in two cases in that study the highest setting made the answers worse. Use the setting the job needs. Leave it low for everything else. ## 6. Let work that can wait, wait Makers list work that can wait up to a day at half price, and text a model has already read at a tenth of the price or less. A weekly report or a nightly clean-up of records doesn't need an answer in two seconds. Ask whoever set up your tools whether they use these discounts. In records from more than a thousand engineering teams, 28% of the calls that could have used the second one did. ## 7. Give each role a default People do match the model to the task, but faintly. Among paying users of one assistant, about half of all use ran on the top model whatever the task was. A default per role makes the choice deliberate: a lighter model for the team that sorts and replies, the strongest for the people whose tasks need it. Leave one door open. A person facing one hard task should be able to pick a stronger model for it. They're the one who can see the task. ## 8. Read the bill by who spent it A seat of an assistant costs about USD 20 to 30 a person a month, and how far it goes depends on the model its user picks. One maker's own estimate is 5 to 45 messages in five hours on its strongest model and 350 to 3,000 on its lightest. The heavy work is moving off the flat fee and onto a meter. So look, once a month, at who spent what and on which model. Look again whenever your supplier releases a new model. The price of a given level of capability has fallen about thirteenfold a year, and that fall goes only to the buyer who moves. The newest model at the top of the list costs more to run each year. ## What we'd wait on Some products offer to choose a model for every single request. The evidence for that is thinner than its headlines. In the largest fair comparison so far, the best methods saved up to 31.7% and several saved nothing, and how well they work depends on traffic they've already seen. A rule you set by kind of work gets you most of what's been measured, and you can read it. ## How SalexHub does it SalexHub sorts its own AI work into five kinds: conversations and the Agent, long tasks the Agent runs on its own, background work, reading emails, meetings and notes, and images. By default it picks a fitting model for each kind. You can choose the model for the Agent and for background work yourself, set which models a team or a person may run, and see spending by model and by who spent it. In the chat, a person can still pick the model for one task. It's all on the [AI Gateway](https://salexhub.com/platform/ai-gateway) page. We haven't measured what an organization saves this way, so we won't quote you a number. --- The same page for people: https://salexhub.com/blog/choose-an-ai-model-for-each-kind-of-work. Index of this site for agents: https://salexhub.com/llms.txt