Claude 5.5: Haiku, Sonnet or Opus, which one to pick?

Anthropic released the 5.5 family in three steps: Opus 5.5 on 22 September 2026, Sonnet 5.5 on 28 September, Haiku 5.5 on 7 October. The API identifiers are claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5. If you only read this paragraph: Sonnet 5.5 for most work, Opus 5.5 when the task is really hard, Haiku 5.5 as soon as you have to process a lot of text on a small budget.

The figures come from Anthropic’s announcements, picked up by specialist press, and from an independent index published by Artificial Analysis. I haven’t reproduced them myself. I cite them with their sources, and I flag when a value is missing or contradicted.

The three models at a glance

	Haiku 5.5	Sonnet 5.5	Opus 5.5
Release	7 October 2026	28 September 2026	22 September 2026
API identifier	claude-haiku-5-5	claude-sonnet-5-5	claude-opus-5-5
Input / output, per million tokens	$0.10 / $0.50	$2 / $10	$4 / $20
Context window	Not confirmed (see below)	1 million tokens	1 million tokens
Maximum output	Not confirmed	128,000 tokens	128,000 tokens
Positioning	The fastest and cheapest	The default setting	The top of the 5.5 family

Haiku 5.5’s prices apply to requests up to 100,000 tokens. Beyond that, the price rises to $0.50 and $2.50.

The price, once the bill is real

The price per million tokens doesn’t give you the final figure. Three things shift it: the size of your requests, caching and batch mode.

Take an example use case: one million input tokens and 200,000 output tokens a month. At standard prices, Haiku 5.5 costs about $0.20, Sonnet 5.5 about $4, Opus 5.5 about $8. This calculation doesn’t measure anything, it just multiplies published prices. Your real figure depends on how long your prompts and answers are.

Caching sharply reduces the cost of re-reading the same long context: one read costs $0.01 per million on Haiku 5.5, and $0.20 on Opus 5.5. Batch mode roughly halves the bill. On Sonnet 5.5, it drops to $1 for input and $5 for output.

Opus 5.5 costs exactly twice as much as Sonnet 5.5. Haiku 5.5 costs twenty times less than Sonnet 5.5 on short requests. Opus 5.5 is 20% cheaper than Opus 5, a price drop that isn’t trivial.

The benchmarks, with their conditions

Anthropic publishes scores from its own tests, and effort settings aren’t always identical from one measurement to another. The table reproduces the values quoted in the detailed sheets. Read them as orders of magnitude.

Test	Haiku 5.5	Sonnet 5.5	Opus 5.5
SWE-bench Pro (fixing bugs in real repositories)	Not cited	81.3%	89.9%
Terminal-Bench 4.0 (command-line tasks)	39.2%	70.6%	66.4%
Humanity’s Last Exam, with tools	57.4%	64.5%	67.7%
GDPval-AA v2.1, office tasks (Elo score)	1,620	1,844	1,846
Artificial Analysis intelligence index, maximum effort	Not cited	56.0	57.6

Two readings stand out. On multi-file bug fixing, Opus 5.5 keeps a clear lead: 8.6 points on SWE-bench Pro. On office tasks, by contrast, Sonnet 5.5 and Opus 5.5 are level, at half the price.

The table also contains a contradiction. On Terminal-Bench, Codersera’s sheet puts Sonnet 5.5 ahead of Opus 5.5, while other sheets make Opus the best score. The gap probably comes from the effort level chosen for each measurement. I draw no conclusion from it.

Which model for which use

Haiku 5.5: volume and latency

Haiku 5.5 is for summaries, classification, information extraction and live customer support. Anthropic also presents it as a sub-agent alongside Opus 5.5 and Sonnet 5.5 on coding tasks: the heavy model orchestrates, the small one runs the repetitive jobs.

Test it on twenty of your own cases before generalising. If the quality holds up, the bill drops fast.

Sonnet 5.5: the default setting

For an agent, a coding assistant or document production, Sonnet 5.5 is the logical starting point. Anthropic aims it at bug fixes, documents, presentations and spreadsheets. According to Codersera’s sheet, it works out cheaper per finished task than Opus 5.5 as long as you stay at a high or lower effort level.

Opus 5.5: when a mistake is expensive

Opus 5.5 is worth it for refactorings that touch many files, for long agentic tasks, and for pure knowledge questions that don’t need external search. It costs twice as much as Sonnet. The maximum effort level raises the bill further, with no guaranteed gain on every task: keep it only if you’ve measured the benefit.

The effort setting, often overlooked

Haiku 5.5 is the first model in its category with an effort setting. It lets you choose between cost and reasoning quality. Opus 5.5 always reasons. Effort is the only lever, not an on/off switch. Its default effort is “medium”, where Opus 5 started at “high”.

In practice, start low for classification, higher for code, and measure. A setting that’s too high doesn’t always make the answer better, it just makes it more expensive.

Frequently asked questions

Should I move from Sonnet 5 to Sonnet 5.5?

The price stays the same, and the announced gains are significant. However, Codersera’s sheet flags API changes that break compatibility. Plan tests before switching production.

Does Haiku 5.5 replace Haiku 4.5?

Haiku 4.5 costs $1 for input and $5 for output. Anthropic announces a workload about 75% cheaper on Haiku 5.5, taking request size into account. That’s the publisher’s estimate, to be checked against your own usage.

What is Haiku 5.5’s context window?

I couldn’t find it confirmed. One source gives 1 million tokens, the others don’t mention it. Check Anthropic’s models page before building anything on it.

Choosing rarely comes down to a spec sheet. Take ten real tasks, run them on the three models at the same effort, then compare the results and the bill. It takes an afternoon, and it tells you more than any table, this one included.