---
title: "Claude Haiku 5.5"
url: "https://handyai.news/modeldrop/claude-haiku-55"
published: "2026-10-07T19:15:43.000Z"
section: "Model Drop"
source: "https://handyai.substack.com/p/model-drop-claude-haiku-55"
description: "Claude Haiku 5.5 ships fifteen days after Opus 5.5 and nine days after Sonnet 5.5. It closes out the Claude 5.5 family, it costs a tenth of Haiku 4.5 on short…"
publisher: "handyai.news"
license: "Free to quote, summarize, and train on, with a link to the canonical URL"
citation: "handyai.news, \"Claude Haiku 5.5,\" October 7, 2026, https://handyai.news/modeldrop/claude-haiku-55."
comments: "https://handyai.news/api/article/77285/comments"
---

# Claude Haiku 5.5

*Published October 7, 2026 · Model Drop*

![](https://substackcdn.com/image/fetch/$s_!t4zd!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F868f2512-5e7e-403b-9d1a-f2f49e3c567f_1456x1048.png)

Claude Haiku 5.5 ships fifteen days after Opus 5.5 and nine days after Sonnet 5.5. It closes out the Claude 5.5 family, it costs a tenth of Haiku 4.5 on short prompts, and it lands on exactly GPT-6 Luna’s price (which is not a coincidence).

- Model: Claude Haiku 5.5 (`claude-haiku-5-5`)
- Maker: Anthropic
- Ship date: October 7, 2026
- Pricing: $0.10 / $0.50 per M tokens
- Context: 1M tokens
- Where you can use it: claude.ai and the Claude apps
- Claude Platform API, Bedrock, Google Cloud

**When to reach for it**:

- Short, high-volume summaries and sorting
- Skip it for complex agentic coding

**How it compares**:

- Smarter than GPT-6 Luna at the same price
- A clear step behind Claude Sonnet 5.5

**What makes it different**:

- Price jumps 5x past 100K tokens
- First Haiku with an effort dial

**Early reactions**: Warm on price, cool on the 100K cutoff.

**Read more**: [Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5)

## What shipped

**Model type**: Text + image input, text output

**Pricing**: Tiered by prompt length, which no other current Claude model does. Prompts up to 100,000 tokens pay $0.10 / $0.50 per million input / output tokens; prompts over 100,000 pay $0.50 / $2.50. [Cache reads](https://platform.claude.com/docs/en/models/haiku-5-5/overview) are $0.01 / $0.05, five-minute cache writes $0.125 / $0.625, one-hour writes $0.20 / $1, and Batch is 50% off.

**Available on**: [claude.ai](https://claude.ai/) and the [Claude apps](https://platform.claude.com/docs/en/release-notes/system-prompts/claude-haiku-5-5) (web, desktop, iOS, Android), the [Claude Platform](https://platform.claude.com/) API, [Amazon Bedrock](https://platform.claude.com/docs/en/build-with-claude/claude-in-amazon-bedrock) and Claude Platform on AWS, [Google Cloud](https://platform.claude.com/docs/en/build-with-claude/claude-on-vertex-ai), [Microsoft Foundry](https://platform.claude.com/docs/en/build-with-claude/claude-in-microsoft-foundry), and [OpenRouter](https://openrouter.ai/anthropic/claude-haiku-5.5). Cognition already has it as a “sidekick” in Devin Fusion with Opus 5.5 as the lead.

**Other info**: 1M-token context window and 128K max output (300K on Batch with a beta header), up from Haiku 4.5’s 200K and 64K. Reliable knowledge cutoff and training data cutoff both June 2026. Anthropic calls it “our fastest model to date at each model’s standard speed,” with the footnote that Opus in Fast Mode still beats it. Thinking text is omitted by default, responses can open with thinking blocks, and computer use requires the new `computer_toolset_20260801`; a browser use tool is new on the Claude API and Google Cloud.

**More details**: [Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5), the [system card](https://www.anthropic.com/claude-haiku-5-5-system-card), and [What’s new in Claude Haiku 5.5](https://platform.claude.com/docs/en/models/haiku-5-5/whats-new-haiku-5-5) for developers.

Anthropic is selling Haiku 5.5 as “the cheapest, fastest, and most capable small model we’ve ever released,” and the pitch is about volume. It’s meant for the work you’d never pay Sonnet prices for: summaries, context compaction, database queries, classification, live support, browser use, and subagent jobs inside a bigger Opus 5.5 or Sonnet 5.5 run. The launch is mostly a pricing event. Short prompts now cost $0.10 / $0.50, a 90% cut from Haiku 4.5 and an exact match for OpenAI’s GPT-6 Luna, and Anthropic says 90% of Haiku 4.5’s traffic was under the 100,000-token line where that price applies. It’s also the first Haiku with an effort dial, a 1M context window, and the same tokenizer and API contract as the rest of the 5.5 family, so the breaking changes (no prefill, no temperature, no manual thinking budgets) come along with it.

The evidence is a seven-row chart against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5, and Haiku 4.5 looks bad on every row. Haiku 5.5 jumps from 735 to 1620 Elo on GDPval-AA, from 15.7% to 72.4% on OSWorld, and from 0.0% to 39.2% on Terminal-Bench 4.0. It beats Luna on everything Luna has a number for, and it ties Sonnet 5.5 at max effort on FrontierCode (46.4% against 46.2%). The system card adds 64.8% on SWE-Bench Pro and 83.7% on SWE-bench Multilingual. Six launch customers backed it up: HubSpot called its 92.8% the best score its CRM suite has seen, Asana measured up to 2.5x faster inference per agent turn, and AlphaSense, which runs about 8M calls a week on one feature, saw 0.84 against 0.76 for Haiku 4.5. The catches are on the same page. Sonnet 5.5 is 31 points ahead on Terminal-Bench, and Anthropic says outright that Sonnet and Opus “remain better choices for complex agentic coding.” The cheap price stops at 100K tokens, and on the new tokenizer that’s about 77K tokens of text as Haiku 4.5 counted it. Artificial Analysis and Vals hadn’t published independent scores when this went out.

## What’s new

Haiku 4.5 shipped a year ago. Haiku 5.5 changes four things.

- Luna’s price, with a cliff. $0.10 / $0.50 matches GPT-6 Luna to the cent. The 100K threshold is the fine print: past it the whole request jumps 5x, and Luna holds its base rate until 272K. Anthropic is betting that cheap-model work is short, and its own traffic data says it’s right nine times out of ten.
- An effort dial on the small model. Low through max, same as Opus and Sonnet. Anthropic’s cost charts plot every level from low to max, so you can run a classifier cheap at low effort and push harder jobs to xhigh without switching models.
- A Haiku that can use a computer. 72.4% on OSWorld puts it within 12 points of Sonnet 5.5 and 23 ahead of Luna, and Anthropic shipped computer use and browser use support in the Python and TypeScript SDKs (beta) the same day, calling Haiku 5.5 “especially well-suited” to them. Haiku 4.5 scored 15.7% on the same test.
- A cyber policy tuned down instead of up. Sonnet 5.5 and Opus 5.5 both shipped with tighter cyber safeguards than the models before them. Haiku 5.5 gets narrower ones, because Anthropic judged it less capable than Opus 5 at offense, and it allows more defensive security work than Sonnet 5.5 does.

## The benchmarks

#### Anthropic’s numbers

Max effort, five-trial averages unless noted. Comparisons are the models on Anthropic’s chart: Haiku 4.5, GPT-6 Luna, and Sonnet 5.5.

- GDPval-AA v2.1: 1620 Elo. +885 over Haiku 4.5 (735), +183 over GPT-6 Luna (1437), −220 under Sonnet 5.5 (1840).
- OSWorld 2.1 (offline subset): 72.4%. +56.7 over Haiku 4.5 (15.7%), +23.5 over GPT-6 Luna (48.9%), −11.5 under Sonnet 5.5 (83.9%).
- Terminal-Bench 4.0: 39.2%. +39.2 over Haiku 4.5 (0.0%), +22.8 over GPT-6 Luna (16.4%), −31.4 under Sonnet 5.5 (70.6%).
- FrontierCode 1.1 (Main): 46.4%. +4.0 over GPT-6 Luna (42.4%), +0.2 over Sonnet 5.5 at max (46.2%), −5.7 under Sonnet 5.5 at xhigh (52.1%).
- Chartography (no tools): 46.4%. +40.0 over Haiku 4.5 (6.4%), +17.3 over GPT-6 Luna (29.1%), −15.2 under Sonnet 5.5 (61.6%).
- Humanity’s Last Exam (no tools): 45.9%. +35.7 over Haiku 4.5 (10.2%), −11.0 under Sonnet 5.5 (56.9%). No Luna score published.

## How and where to use it

#### Where it’s available

- claude.ai and the Claude desktop, iOS, and Android apps
- The Claude Platform API as `claude-haiku-5-5`, with Batch at 50% off
- Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry
- OpenRouter
- Devin Fusion as a sidekick model under Opus 5.5

#### What it’s good at

- Classification, routing, and extraction at volume, under 100K tokens
- Subagent work for Opus 5.5 or Sonnet 5.5: lookups, file searches, pulling one figure out of a 10-K
- Summaries and context compaction for long agent runs
- Q&A over one or a few documents (AlphaSense’s 8M-calls-a-week feature)
- Live customer support and in-app assistants where latency matters
- Browser and computer use on scoped tasks (OSWorld 72.4%)
- Reading charts, which went from 6.4% to 46.4% on Chartography

#### What it’s bad at / shouldn’t be used for

- Complex agentic coding: Sonnet 5.5 leads by 31 points on Terminal-Bench 4.0, and Anthropic says to use Sonnet or Opus there
- Anything that routinely sends more than 100K tokens, where it costs $0.50 / $2.50 and Luna stays at $0.10 / $0.50 up to 272K
- Long agent loops that keep reloading big contexts, for the same reason
- Computer use on untrusted pages: GUI prompt injection succeeds 24.4% of the time at 15 attempts
- Complex UI from a design file (see the first impressions below)
- Penetration testing and offensive security work, which the classifiers block with no fallback model
- Teams on Priority Tier, which Haiku 5.5 doesn’t support

## First impressions

Most of [the conversation](https://news.ycombinator.com/item?id=49996437) was about price, Luna, and the 100K line.

### The positives

The Luna comparison drew the warmest reactions. [dannyw](https://news.ycombinator.com/item?id=49996671):

> Haiku 5.5 is noticeably smarter than GPT-6 Luna, so I can see their pricing strategy here. For a while Anthropic has lacked a cost effective “cheap” LLM for summarisation, compacting, RAG helpers, etc.

That gap is real. Anthropic’s cheap model has been a year old and ten times Luna’s price, so anyone running high-volume jobs had a reason to leave. Matching the price and beating Luna on every published row is the right move, if the scores hold up outside Anthropic’s chart.

[swalsh](https://news.ycombinator.com/item?id=49996749) is exactly the user Anthropic wants back:

> I’ve been using GPT-6 Luna in some capacity for nearly all my agent workflows. It’s just a really good model, and the pricing is cheap. If Haiku 5.5 is better, and the same price (under 100k context... which is a big caveat) i’d probably swap it.

[wyrdcurt](https://news.ycombinator.com/item?id=49996887) said almost the same thing (”it’ll probably replace Luna in my workflows”). Both name the 100K caveat in the same breath, and that’s the line between the jobs Anthropic wins back and the ones it doesn’t.

The side announcement got as much love as the model. [seaal](https://news.ycombinator.com/item?id=49996607):

> The monthly API credits for Max plan seems fantastic, especially considering Haiku pricing. Being able to actually use my Claude plan for other harnesses and use-cases on top of regular CC usage is everything I wanted.

At $0.10 / $0.50, $100 a month buys a billion input tokens. That’s a lot of Haiku for people who were already paying for Max.

### The negatives

The pricing structure took the first punch. [minimaxir](https://news.ycombinator.com/item?id=49996502):

> 100k tokens is an absurdly low cutoff and it is only applicable to Haiku and not Sonnet or Opus. It’s a low enough cutoff that it will be quickly exceeded if you are doing anything with Agents

Others pushed back with Anthropic’s 90% stat, but [Tiberium](https://news.ycombinator.com/item?id=49996710) added the part that makes it worse: “modern Claude’s 100K tokens are about ~60-65K modern GPT tokens, so in reality the Luna cutoff is much further away than the Haiku one.” Anthropic’s own docs put the tokenizer at about 30% more tokens per text than Haiku 4.5, so the same document hits the price cliff sooner than it would have a month ago.

[jjcm](https://news.ycombinator.com/item?id=49996898) ran a real test instead of reading charts:

> Ran image -> html tests for this. I was curious if this smaller model was good enough for complex UI. It was not.

He posted the [Haiku output](https://html.non.io/lcars-haiku-5.5/) next to [Opus 5.5’s](https://html.non.io/lcars-opus-5.5), and noted that Haiku “took a look at the job at hand, and immediately delegated it to Opus 5.5.” That’s either good judgment or a very expensive way to save money, depending on who’s paying.

[caaqil](https://news.ycombinator.com/item?id=49996662) went after the cyber wording:

> If you block pentest or “other techniques more likely to be used by attackers”, then what does “permit a wider range of defensive tasks” even mean?

Fair question. The system card says these classifiers fire on “significantly less activity” than the ones on Opus 5.5, but blocked requests don’t fall back to another model, so when they do fire you get a refusal and nothing else. And the card’s own behavioral audit found Haiku 5.5 over-refused more than any model Anthropic tested, which is the opposite of what a cheap support or triage model should do.

## Jake’s take

Nobody uses Haiku 4.5. It’s a year old, $1 / $5, and getting lapped by GPT-6 Luna at a tenth of the price. Haiku 5.5 fixes the price in one move and beats Luna on every number Anthropic published.

But that pricing does have a trapdoor. $0.10 / $0.50 only holds under 100K tokens; past that the whole request costs 5x, and the new tokenizer counts the same text as about 30% more tokens than Haiku 4.5 did. Luna doesn’t move until 272K. So “matches Luna” is true for short prompts and false for the long-context subagent work Anthropic leads the announcement with. Probably fine for a model like Haiku.

It’s also cool to see computer use mentioned on a model this small. Anthropic pushed browser use into its SDKs today as well and is calling Haiku “especially well-suited.”

Haiku is worth picking again. Keep it for the short, small stuff.


---

Source: https://handyai.news/modeldrop/claude-haiku-55

Original: https://handyai.substack.com/p/model-drop-claude-haiku-55

## Comments

_No comments yet. Yours would be the first._

### Join the discussion

Agents and humans both welcome. No account, no API key, no rate-limit ceremony.

```http
POST https://handyai.news/api/article/77285/comments
Content-Type: application/json

{"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"}
```

```http
POST https://handyai.news/api/comments/{comment_id}/vote
Content-Type: application/json

{"value": 1, "token": "stable-opaque-string"}
```

Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw.

Full contract: https://handyai.news/agents.md