Match the model to the job. A cheap, fast model handles routine work such as rewriting, reformatting and short summaries, and a stronger model earns its price when the task is hard or a mistake would be expensive. Start cheap and move up only when the results are visibly weak.
Models differ on the three things you'll feel: how well they cope with difficult work, how quickly they reply, and how much each reply costs. This guide covers the terms you'll meet on a price list and a simple way to decide.
What is an AI model?
An AI model is the system, trained on vast amounts of text (and often images and audio too), that does the thinking when you send a message. When people say "AI", they usually mean a model such as GPT, Claude or Gemini.
On its own, a model answers and then forgets. It doesn't keep notes about you, run a task at 7am or use tools. The software that adds memory, schedules and access is usually called an agent engine or harness, and hosted vs self-hosted AI agents explains how the parts fit together. For the difference between asking a model questions and working with an agent, see AI usage vs an AI agent.
That matters when you compare prices. Two services can run the same model and feel nothing alike, because one wraps it in a proper agent and the other gives you a chat box.
What do AI credits pay for?
Tokens. Every message you send and every reply you get is measured in tokens, small chunks of text that are often part of a word, and providers bill by the token. How much text a token covers depends on the model and the language, as Anthropic's glossary explains. A short email might be a few hundred tokens; a long report can run to tens of thousands. Replies usually cost more per token than what you send, which you can check on any provider's price list, such as Anthropic's pricing page.
So a summary of a two-page letter costs very little, while asking an agent to work through a long report and produce a detailed plan costs noticeably more.
You'll usually pay in one of two ways.
Prepaid credits. You buy a balance and each message draws it down. You can see what you're spending and there's no surprise bill at the end of the month. The trade-off is remembering to top up.
A subscription you already have. Some platforms let you sign in with a ChatGPT or Claude plan so your usage runs against it instead of credits. That's good value if you already pay for one and the provider allows it, and the two big providers take different lines:
- OpenAI lets Go, Plus and Pro subscribers use their ChatGPT plan in participating apps through "Sign in with ChatGPT". Usage counts against the plan's limits, and you can set a weekly limit for each app.
- Anthropic limits Claude subscription sign-in to its own apps, such as Claude Code, including where a platform runs an unmodified copy of Claude Code for you. Other third-party apps can't offer Claude login or route requests through Free, Pro or Max plans; they have to use an API key.
These rules change. Check the provider's current terms before you rely on a route, and trust written confirmation over a forum thread.
How do quality, speed and cost trade off?
You can't have the best of all three at once, so decide which one the task needs most.
Quality. Stronger models cope better with ambiguity, long documents and reasoning across several steps. On easy tasks you often can't see the difference; on hard ones it's obvious.
Speed. Bigger models tend to reply more slowly: a second or two on a quick question, up to a minute on a long piece of writing. If you're waiting mid-conversation, speed can matter more than a small gain in quality.
Cost. You roughly pay for capability. A provider's premium model can cost many times more per token than its cheapest one, so the same request can vary a lot in price.
Which jobs need a stronger model?
The ones where a mistake is expensive, or where the answer depends on holding a lot in mind at once. Most work splits roughly like this.
A cheap, fast model is usually enough for:
- rewriting and shortening text
- turning notes into a list
- translating routine messages
- quick questions where you'll check the answer anyway
- summarising something short
Move to a stronger model for:
- contracts, figures and documents that go to clients
- long documents where the answer depends on several sections together
- planning with lots of constraints
- anything you'd normally take to a senior colleague
Pay for the very top model only when you can say what it buys you on this particular task. Starting cheap and stepping up when a result is weak shows you where the line sits for your work; starting at the top and working down costs more and tells you less. Most of the jobs in everyday tasks for an AI agent run happily on a cheap model.
Does the context window matter?
Only for long material. A context window is how much text a model can consider at once: the current conversation plus anything you've pasted into it. Once it fills up, older material gets dropped or compressed.
You'll notice it when you paste a long report or keep one thread going for weeks. If your agent loses track of how a long conversation started, that's usually why, and nothing is broken.
Persistent memory is a separate feature: the notes an agent keeps about you between conversations. How agent memory works covers the difference.
Keeping the bill predictable
If you buy credits, a few settings do most of the work:
- A daily cap. A limit you rarely hit costs nothing and stops a runaway task from draining your balance.
- A low-balance alert. Running out halfway through a job is avoidable.
- A cheaper model for routine work. Most waste comes from sending everyday tasks to a premium model, far more than from sending too many messages.
With Able°, you choose a model by cost tier and can set both a daily cap and a low-credit alert. On a subscription sign-in, look at the provider's usage page now and then, so you notice if a workflow starts using far more than you expected.
Testing models on your own work
A price list tells you what a model costs, and nothing about whether it's good at your kind of work. Benchmarks help a little, but they measure tasks that may look nothing like yours.
Take three real tasks from this week, run each one on a cheap model and a stronger one, and compare the results yourself. Half an hour of that tells you more than any comparison table.
If you send only a handful of messages a day, the price gap between models is small enough to ignore: pick a sensible default, set a cap and move on. The choice deserves your time when usage is heavy, when the output goes to clients, or when a wrong answer would cost you money.
Common questions
Which AI model is best?
No single model is best at everything. The strongest models handle hard reasoning and long documents better, but they're slower and cost more, while cheaper ones do routine writing and summaries perfectly well. The best model for you is the cheapest one that does your task to the standard you need, and the only way to find it is to test on your own work.
Can I switch AI models later?
Usually, yes. In most agents, memory and skills are kept by the agent software rather than inside the model, so they stay put when you change model. Expect the writing style and the quality on hard tasks to shift a little, and rerun a couple of your regular jobs to check the new model still meets your standard.
Why do some AI models cost more than others?
Bigger, more capable models need more computing power for every token they read and write, so providers charge more for them. The extra money buys better handling of difficult reasoning, long documents and subtle instructions. For a job like rewriting a message, that extra capability often makes no visible difference.
Can I use my ChatGPT subscription in other apps?
Sometimes. OpenAI allows it in apps that offer "Sign in with ChatGPT", with usage counted against your plan's normal limits, but not every app takes part. Claude plans are more restricted and work only in Anthropic's own apps, such as Claude Code. Check the current terms, because both providers update them.