- Makers split their offers into three classes: flagship, balanced and fast with a low price.
- For classification, data extraction and routing the fast class is usually enough, and for hard reasoning and long agentic tasks reach for the flagship.
- Pick the cheapest model that passes your own test.
The most common mistake in choosing a model is picking the “best” one from a list. The best according to the maker on its chart usually means the most expensive and slowest, and that is not the right answer for most tasks in an application. The right answer is: the cheapest model that passes your test. To find it you need to know how makers split their offer and what they themselves say about it.
Three classes of models
The offers of the largest makers fall into three classes, and Anthropic and OpenAI describe them similarly. The flagship class is meant for demanding reasoning and long agentic tasks. The balanced class combines capability with cost and is usually a sensible starting point. The fast and cheap class serves high-volume, latency-sensitive tasks such as classification, data extraction and routing of requests. Anthropic’s page advises starting with Opus 5.5 for most workloads and reaching for Fable 5.1 when your own evals at higher effort still fall short.
| Flagship | Balanced | Fast and cheap | |
|---|---|---|---|
| Anthropic | Fable 5.1: demanding reasoning, long agentic tasks | Opus 5.5 and Sonnet 5.5: coding, knowledge work, speed and intelligence | Haiku 5.5: classification, extraction, routing |
| OpenAI | GPT-6 Astra: complex reasoning and coding | GPT-6.1 Sol: a balance of intelligence and cost | GPT-6 Luna: cost-sensitive, high-volume tasks |
| Typical cost | Highest | Medium | Lowest |
| Typical speed | Slower | Moderate | Fastest |
The split is not rigid. Makers also offer specialized versions and reasoning effort levels, which let you trade time and tokens for accuracy within one model. As a result the choice is not a single point but a combination: the model, the effort level and the prompt. So when comparing, record all three elements and the token usage, because the same model at a higher effort level can cost several times more than at a lower one. It happens that a cheaper class with higher reasoning effort passes a test for less than a more expensive class on default settings, and you will not see that if you compare only model names.
How to match a class to a task
Start with the kind of work. Tasks with clear rules and short input, such as assigning a ticket to a category, extracting dates and amounts from an invoice, or initial filtering, usually fit the fast class. Generative tasks of medium difficulty, such as content drafts, summaries and customer replies, are the territory of the balanced class. Tasks where an error is costly, or that need planning across many steps, for example a coding agent working for hours, justify the flagship. This is a rule of thumb, not a law: check it with a test.
- 01TaskDescribe it and the criterion for passing the test.
- 02SetCollect 20–50 cases, including hard ones.
- 03MiddleRun the balanced class.
- 04DownTry the fast class.
- 05UpThe flagship only if the middle fails.
- 06ChoiceThe cheapest model that passes.
Testing in both directions matters. Merely checking that a flagship model does great proves nothing, because a cheaper one would probably have done fine too. Only comparing the result with the cost shows how much you pay for extra quality and whether you need it. How to build such a set is in the article How to evaluate an agent before you trust it.
Routing: several models in one application
The write-up on effective agents describes a routing pattern: the input is classified and directed to a specialized next step. In cost terms it means a cheap model decides whether a request is simple (then it answers itself) or hard (then it hands it to a stronger one). Such a setup often lowers the bill without a drop in quality, provided the classification is well tested and the stronger model really only gets hard cases. It is also worth considering the cost of a task in tokens, to know how much you save.
Pitfalls of choosing
The first pitfall is choosing by name and brand, without a test. The second is ignoring latency: a model that answers twice as slowly can spoil the user’s experience. The third is forgetting about retirement: keep the model ID in configuration and have a test that shows regressions after a version change. The fourth is blind trust in rankings; how to read makers’ pages and comparisons is explained by the text Frontier models: how to read vendor model pages. You will find more materials in the Frontier models hub.
Last checked: 20 October 2026. Class descriptions come from makers’ pages of that day and will change.