Anthropic names its models by tier, not by number. Each tier is a different trade-off between capability and price, and each generation (4.5, 4.6, 5, 5.1…) refreshes the ladder. In the app you pick from a model menu; in the API you pass a model id string. Either way, the choice is where most of your money goes.
The four tiers
| Tier | Current model | Best for | Context | Rule of thumb |
|---|---|---|---|---|
| Haiku | Claude Haiku 4.5 | High-volume, latency-sensitive work: classification, extraction, chat at scale | 200K | Cheapest — try it first for anything repetitive |
| Sonnet | Claude Sonnet 5 | The everyday workhorse: writing, coding, analysis, most agent steps | 1M | Default for 80% of tasks |
| Opus | Claude Opus 5 | Hard reasoning, long autonomous coding sessions, research synthesis | 1M | When Sonnet's answer is not good enough |
| Fable | Claude Fable 5.1 | Anthropic's most capable widely released model; the hardest problems where correctness beats cost | 1M | Twice the Opus rate — earn it on a benchmark first |
Older generations — Sonnet 4.6, Opus 4.8, Opus 4.7 — remain available on the API at their own rates and are sometimes cheaper. The catalogue lists them all on the Models & Providers page. For new work, start on the current generation: the newer models at lower effort settings usually beat the older ones at full effort, and you keep one price list to manage.
Context windows
Current Sonnet, Opus and Fable models accept up to a million tokens — roughly 750,000 words — in one request. Haiku 4.5 accepts 200,000. That is enough for entire codebases or hundreds of pages, and it is the feature that makes Claude popular for document-heavy work.
It is also the feature that quietly drains API budgets. A long conversation re-sends its whole history on every turn, so the fortieth message in a chat with a pasted 100-page PDF bills that PDF for the fortieth time. In the app this shows up as hitting your usage limit faster; on the API it shows up on the invoice. Lesson 8 covers prompt caching, which is the fix.
Thinking and effort
Current Claude models can *think* before answering — reason through a problem in private before writing the response. In the app this is the extended thinking toggle. On the API it is adaptive by default on the newest models and you steer it with an effort setting from low to max. Higher effort means better answers on hard problems and more output tokens billed. For routine tasks, lower effort on a current model is both cheaper and often better than an older model at full effort.
Knowledge check
You need to tag 50,000 support tickets by topic. Which model should you try first?