Claude 5.5: A Field Guide for Builders
Four Claude models, one family. What each is for, what it costs, how caching changes the maths, and how to pick without overspending.
October 19, 2026 · AI · Guide
Anthropic's Claude family now has four current models, and the newest, Haiku 5.5, arrived on 7 October 2026. This guide covers what each one is for, what it costs, and how to avoid paying for more model than a job needs.
Every number here comes from Anthropic's own pricing and model pages, checked on 19 October 2026. Prices change, so the sources are listed at the end.
The line-up
Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5 all take text and image input, support tool use, and offer a 1M-token context window with up to 128K tokens of output. They differ in how fast they answer, how much they think, and what they cost. Pick one to see its details:
Anthropic's own advice is to start with Opus 5.5 for most workloads, and to reach for Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your own tests on Opus 5.5 at higher effort still fall short.
What each one costs
List prices per million tokens. Anthropic charges separately for the text you send in, the text that comes back, and any input served from a cache.
Two details are easy to miss. The three larger models charge the same rate across the whole 1M-token window. Haiku 5.5 is the exception: it is priced by prompt size, so a request over 100,000 tokens moves to a higher band ($0.50 in, $2.50 out instead of $0.10 and $0.50).
Prices keep falling
The newest generation is cheaper than the one before it, sometimes by a lot. Anthropic says Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average, and that halving Sonnet 5.5's cache-read price makes it around 20% cheaper on most agentic work. The list prices behind that:
Effort, instead of guesswork
Every current Claude model uses adaptive thinking: the model decides how much to think, steered by an effort setting you control. Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, which means the cost-versus-intelligence dial now exists across the whole family.
| Model | Default effort | Latency |
|---|---|---|
| Fable 5.1 | high | Slower |
| Opus 5.5 | medium | Moderate |
| Sonnet 5.5 | high | Fast |
| Haiku 5.5 | medium | Fastest |
Effort and model choice are two separate dials. It is worth testing both, because a cheaper model at higher effort and a pricier model at lower effort can land in very different places on the same task.
Caching is the biggest lever
When part of a prompt repeats, such as a long system prompt, a document, or a codebase, a cache hit costs a fraction of normal input. How small a fraction depends on the model:
Anthropic's own worked example shows the effect: a one-hour coding session on Opus 5.5 using 50,000 input tokens and 15,000 output tokens, run as a Claude Managed Agents session.
That is about 26% less for the same work. The Batch API adds a separate 50% discount on input and output for work that can wait, and the two discounts stack.
Claude Code
Alongside the API, Anthropic ships Claude Code, an agentic coding tool that reads a codebase, edits files, runs commands, and connects to development tools. It runs in the terminal, in VS Code and JetBrains, in a desktop app, and in the browser, and every surface shares the same underlying engine, so a project's instruction files and settings carry across them.
It can also be scripted: input can be piped into it and it can run non-interactively, which makes it usable from CI jobs as well as from a keyboard. Because it runs on Claude models, the trade-offs above between cost, speed, and effort apply to coding work too.
Picking a model
A rough map from the job to the model, based on each vendor's own descriptions of what its models are for. The second card shows the nearest OpenAI equivalent, which the price comparison covers in more detail.
How we would approach it
Three habits cover most of the cost advice in this guide:
- Start at the cheapest tier that could plausibly work, test it on your own tasks, and move up only when the results demand it.
- Re-run your tests when models change. Prices and quality both moved within weeks this autumn, so last quarter's choice may no longer be the right one.
- Keep the repeated part of a prompt stable so it can be cached, and put anything that changes at the end.
If you would rather try several models in one place before committing, Crooli AI brings them into one workspace.