# Claude Haiku 5.5: Anthropic's small, cheap model for high volume and subagents

> What Claude Haiku 5.5 is, Anthropic's published benchmarks and how to read them, what it costs (from $0.10/$0.50) and when to pick it. As of November 15, 2026.

*Tags: ai, models, anthropic, low-cost, subagents*

---


> [!info] Quick Definition
> **Claude Haiku 5.5** is the small model in Anthropic's Claude 5.5 family, the cheapest and the one with the lowest comparative latency, released on October 7, 2026. It is built for high-volume, low-latency work such as classifying, extracting data, routing requests and running subagents.


## What is Claude Haiku 5.5 and when did it come out?

It is the budget option in the family, alongside [[claude-opus-5-5|Opus 5.5]], Sonnet 5.5 and [[claude-fable-5-2|Fable 5.2]]. The documentation ranks it as the lowest-latency model in the comparison. It is a clear example of what [[small-vs-large-models|small vs large models]] discusses: lower cost and more speed in exchange for less capability on hard tasks.

| Fact | Value |
|---|---|
| API ID | `claude-haiku-5-5` |
| Release | October 7, 2026 |
| Context and output | 1M tokens of context; up to 128K of output |
| Price | $0.10 input / $0.50 output up to 100,000 prompt tokens; $0.50 / $2.50 above that (per million tokens) |
| Input and output | Text and images in, text out |
| Knowledge cutoff | June 2026 |
| Default effort | `medium` |
| Open or closed | Closed: used through the API and cloud platforms |

## What is it for?

Anthropic describes it for high-volume, cost-sensitive work: summaries, context compaction, database queries and classification. Also as a subagent next to Opus 5.5 and Sonnet 5.5 on coding tasks, and for speed-sensitive tasks such as live customer support and browser use. The documentation adds extraction and routing.

A [[subagents|subagent]] is a secondary agent to which a main agent delegates part of the work. That is why it fits: the large model plans and the small one runs thousands of repetitive steps at low cost.

## Which benchmarks did it publish and how do you read them?

Anthropic reports these scores in the Haiku 5.5 announcement, using its own test harness. They are not an independent measurement.

| Benchmark | Score | Measured by | What it measures, per the source |
|---|---|---|---|
| Terminal-Bench 4.0 | 39.2% | Anthropic | How well it completes complex, multi-step professional tasks in a command-line interface |
| OSWorld 2.1 (offline subset) | 72.4% | Anthropic | How well an agent operates a real computer on long, multi-step tasks |
| Humanity's Last Exam, no tools | 45.9% | Anthropic | Expert-level academic knowledge and reasoning |
| Humanity's Last Exam, with tools | 57.4% | Anthropic | The same, with tools available |
| GDPval-AA v2.1 | 1620 Elo | Anthropic | Real-world professional work across 44 occupations |
| FrontierCode 1.1 (Main) | 46.4% | Anthropic | The page does not describe it |

How to read them:

- **Compare with the previous version, in the same table.** Anthropic reports Terminal-Bench 4.0 at 0.0% for Haiku 4.5 and 39.2% for Haiku 5.5, and OSWorld 2.1 at 15.7% and 72.4%. Same test, same lab.
- **Watch the variants.** Humanity's Last Exam appears with and without tools, and the scores differ by more than 10 points. A number without its variant is not comparable.
- **Do not build rankings against other labs.** Anthropic's page puts other models in its comparisons (for example GPT-6 Luna), but a score from another lab was not measured with the same harness.
- **A small model is measured on your task.** For classifying or extracting, what matters is the hit rate on your data, not a general exam. The guide is in [[ai-benchmarks|how to read AI benchmarks]].

## How much does it cost and how much context does it have?

As of November 15, 2026, Anthropic's documentation lists these prices per million tokens:

| Prompt length | Input | Output | Cache read |
|---|---|---|---|
| Up to 100,000 tokens | $0.10 | $0.50 | $0.01 |
| Over 100,000 tokens | $0.50 | $2.50 | $0.05 |

Batch processing (Batch API) has a 50% discount. The pricing page clarifies that prompt length counts all input tokens, including cache reads and writes, and that each request is priced separately. A 150,000-token prompt pays everything at the higher rate.

Two details when comparing with Haiku 4.5 ($1 / $5): Anthropic says Haiku 5.5 costs about 75% less on average (90% less up to 100,000 prompt tokens and 50% less above that), and that this figure already includes the newer tokenizer using somewhat more tokens per task. Its documentation specifies that the same text counts as roughly 30% more tokens than on Haiku 4.5. To understand the units, read [[tokens|tokens]], [[context-window|context window]] and [[cost-per-token|cost per token]].

Illustrative example of the math: classifying 10,000 tickets, each with 2,000 input tokens and 300 output tokens, is 20 million input tokens and 3 million output tokens: $2.00 + $1.50 = $3.50 ($1.75 in batch). The same job on Opus 5.5, at $4 / $20, would cost $80 + $60 = $140. Real spend varies with what the model generates.

## When should you pick it and when not?

Pick it when:

- You have thousands of simple, repetitive tasks: classifying, extracting fields, routing or summarizing.
- Latency matters, as in a live assistant.
- You want cheap subagents under a more capable model.

Do not pick it when:

- The task needs hard reasoning or complex code changes: that is where [[claude-opus-5-5|Opus 5.5]] or [[claude-fable-5-2|Fable 5.2]] go.
- A mistake is costly and there is no check afterward.
- Your prompts routinely exceed 100,000 tokens: you will pay the higher rate.

To see options from other labs, check [[current-ai-models|current AI models]].

## Example

Illustrative example: a product team receives 5,000 user comments a month and wants to tag each one as "bug", "feature request" or "question". They take 100 comments, tag them by hand and run the same list through Haiku 5.5. They count the hits and review the misses. If the rate is acceptable, they process the rest in batch and send the doubtful cases to a person.

What to try tomorrow: build a list of 50 cases from your product with the correct label, run it on Haiku 5.5 and on Opus 5.5 with the same prompt, and compare hits and cost. If the difference is small, you save with the cheap model.

## Common mistakes

- Choosing Haiku on price without measuring hits on your own data.
- Sending prompts over 100,000 tokens without noticing that the rate changes.
- Comparing cost with Haiku 4.5 without accounting for the tokenizer counting more tokens.
- Using a general score as a guarantee for a specific task.
- Letting it decide with no human review in high-risk cases.

## Sources

- [Anthropic: introducing Claude Haiku 5.5](https://www.anthropic.com/news/claude-haiku-5-5)
- [Documentation: Claude Haiku 5.5](https://platform.claude.com/docs/en/models/haiku-5-5/overview)
- [Documentation: Claude models overview](https://platform.claude.com/docs/en/about-claude/models/overview)
- [Documentation: pricing](https://platform.claude.com/docs/en/about-claude/pricing)


---

Source: https://www.fernandoux.com/en/wiki/ai/claude-haiku-5-5/
