Choosing the right AI model for the job: tokens, effort and cost

The largest AI model is not the right one for every task. Here is how tokens set the bill, what model size, effort and thinking each control, and how to keep costs in proportion as AI spreads through a company.

Published
Reading time
6 min read
Written by
LekkerLogic

Key takeaways

  • AI is billed by the token, and the largest model costs many times what the smallest does. The model chosen for each task sets most of the bill.
  • Model size, effort and thinking are separate settings. Routine text work belongs on a small, fast model. Analysis that people act on is worth a large one.
  • As use grows, set the model feature by feature, cache what repeats and measure. For client work we use US-based providers and do not use DeepSeek, Qwen or other models from outside the US.

Most people meet AI through a chat window with one model behind it. Building AI into the daily work of a business is different. Tidying a note, sorting an email and briefing a manager on an account are each a separate request to a model, and somebody has to decide which model answers and how hard it should work.

That decision sets the quality of the answer, how long the person waits and what the request costs. The examples below use Anthropic's Claude models, which we build with most often. The same ideas apply to OpenAI.

Tokens: what AI is billed in

A model does not read words. It reads tokens, which are short pieces of text. A common word is one token, a long word is two or three, and a page of writing is several hundred. Providers bill by the token for everything you send (the input) and everything the model writes back (the output).

  • Output costs more than inputOn Claude's current price list, a token the model writes costs five times as much as a token it reads. Long answers are the expensive part.
  • Everything sent is counted every timeInstructions, background documents and the earlier conversation are billed again with each request. The fiftieth message in a long chat costs far more than the first.

One provider, several models

Each provider sells a range of models, from small and fast to large and capable. Anthropic's current range has four tiers.

Table 1The Claude model tiers, with the price per token relative to the smallest.
TierSuited toRelative price
HaikuSmall and fast. Short, well-defined text work at high volume: tidying wording, sorting messages, pulling fields out of a document.1×
SonnetThe everyday middle. Drafting, summarizing and multi-step tasks that need judgment.20×
OpusLarge. Analysis across a lot of material, where a person will act on the conclusion.40×
FableThe most capable. The hardest reasoning and long, multi-step work.80×

Relative prices compare Anthropic's published API list prices for Claude Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.2 at the time of writing. Prices change, so check the current list.

That 80-fold spread is why model selection matters. Take one routine task, reading a short email and filing a two-line summary, and run it 10,000 times a working day across a company. At list prices that comes to about $100 a month on the smallest model and about $7,900 a month on the largest, for a task the smallest model does well.

The example assumes about 2,000 tokens in and 500 out per request and 22 working days a month, before any caching.

The opposite mistake is expensive in another way. A small model asked to analyze a year of sales activity returns a confident, shallow answer, and a manager may act on it.

Effort and thinking: how hard the model works

The model is the first setting. Two more decide how much work goes into each answer, and both affect how long the person waits.

  • EffortA setting from low to maximum that tells the model how thorough to be. Higher effort means more checking, more tokens and a longer wait.
  • ThinkingBefore it answers, a model can reason through the problem privately, like working on scratch paper. That helps on hard questions. Thinking is billed as output, so on a simple task it is wasted money.
  • SpeedSmall models and low effort answer sooner. For a button someone clicks forty times a day, a two-second answer matters more than a brilliant one.

One client build, three different settings

A custom sales platform we built for a client has three AI features close together on screen, each with its own model and its own effort setting.

Table 2Three AI features on one client build, each with its own model and settings.
The jobModelEffort and thinkingWhy
Tidy the wording of a typed noteSmall (Haiku)Low effort, thinking offThe task is narrow: improve the wording, never change the facts. The person is waiting on the button.
Draft presentation slides from a manager's notes and filesMid-size (Sonnet)High effort, thinking onThe model has to plan which source supports which point. The job did not need the largest model.
Brief a manager on a business unit's sales pictureLarge (Opus)Moderate effort, thinking onIt reads across a long history and says what changed and where to look. A person acts on it, so depth is worth paying for.

Swap the first and last rows and both features get worse. The clean-up would be slower and cost many times more for no visible gain. The briefing would miss the patterns a manager needs, and that is the work where an Opus or Fable class model earns its price.

Keeping costs in check as AI use grows

One person trying AI costs little, whatever the model. With fifty people, a dozen features and thousands of requests a day, token use becomes an operating cost like any other and needs the same attention.

  1. Set the model feature by featureA single company-wide default is either too expensive for routine work or too weak for analysis.
  2. Cache what repeatsMost features send the same long instructions with every request. Prompt caching lets the provider keep that repeated part ready for a short time and reuse it at a tenth or less of the normal input price.
  3. Send less, and let software do the arithmeticGive the model the fields a task needs, not the whole record. Do totals and lookups in ordinary code, which is exact and nearly free, and ask the model to explain the finished numbers.
  4. Batch what can waitWork nobody is waiting on, such as overnight summaries, can run as a batch at half the price.
  5. Measure by featureLog token use against each feature and review it monthly. The surprise is usually one feature set a tier too high, or one set of instructions that has quietly grown.

Claude, OpenAI and the models we do not use

None of this is specific to one provider. OpenAI's range follows the same pattern: smaller, cheaper models beside larger reasoning models, billing by the token with output priced above input, a reasoning effort setting, and discounts for cached input and batch work. We build with both and keep the model as a setting, so moving a feature to another provider is a configuration change, not a rebuild.

Start with the work, not the model

Model selection goes wrong when it starts from the model, so we start from the work. When LekkerLogic consults with a team, we sit with the people who do the job and list where the day goes: reading, re-keying, searching, drafting, summarizing and checking. Each task falls into one of three groups.

  • Routine and frequentTidying, sorting, extracting and first-pass replies. A small model at low effort, built into the screen where the work happens, with a person able to correct or undo it.
  • Judgment-heavyAccount reviews, proposals and analysis. A larger model prepares the draft or the briefing, and a person decides what to do with it.
  • Not an AI taskCalculations, rules and lookups. Ordinary software does these exactly, for almost nothing.

In every group the model supplements the person. It takes on the reading and the first drafts, and people keep the decisions and the customer conversations. Matching the model to each task lets that spread across a company without the bill growing faster than the value.

Model names and prices are as published at the time of writing and change often. LekkerLogic is not affiliated with the providers named.

Not sure which model belongs on which task? LekkerLogic works with your team to map the day-to-day work, match each task to the right model and settings, and build the tools with cost controls in place.

Start a conversation
Share LinkedIn (opens in a new tab) Email