---
title: Claude Haiku 5.5: what improved, and when to use Fable, Opus, Sonnet or Haiku
seo_title: Claude Haiku 5.5: what improved and which Claude model to use
date: 2026-10-08
summary: Anthropic released Claude Haiku 5.5 on October 7: 90 percent cheaper than Haiku 4.5 for most prompts, faster, and far stronger on its benchmarks. What improved, when to use Fable, Opus, Sonnet or Haiku, and how one model can orchestrate the others.
description: What improved in Claude Haiku 5.5, when to use Fable, Opus, Sonnet or Haiku, and how Claude Code and the Claude Platform let one model orchestrate the others.
tags: [ai, claude, claude-code, models]
---
Anthropic released Claude Haiku 5.5 on October 7, 2026. It is the small model of the family, it lists at a tenth of Haiku 4.5's price for most prompts, and it is the first Haiku with an adjustable effort setting. This article says what improved, when to use Fable, Opus, Sonnet or Haiku, and whether anything orchestrates the four.

I am writing this a day after the release, so it is a reading of Anthropic's own pages, not a test of the models. Every fact comes from the pages listed at the end, which I read on October 8, and the benchmarks are Anthropic's own. The job in the cost example is made up, and the routing advice is a design to try.

## What is new in Claude Haiku 5.5?

Claude Haiku 5.5, released on October 7, 2026, is what Anthropic calls its cheapest, fastest and most capable small model, and its list price is 90 percent lower than Haiku 4.5's for prompts up to 100,000 tokens. The facts, as Anthropic gives them:

- **Released:** October 7, 2026, on all platforms, including AWS, Google Cloud and Microsoft Azure. The API ID is `claude-haiku-5-5`.
- **Price:** $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, against $1 and $5 for Haiku 4.5, and $0.50 and $2.50 above that. Anthropic says it costs around 75 percent less to run on average.
- **Speed:** Anthropic's fastest model at standard speed, though slower than its Opus models in Fast Mode. Box measured about half the latency of Haiku 4.5.
- **Effort:** the first Haiku with an adjustable effort setting. Adaptive thinking is on by default, and the default effort is medium.
- **Size:** a 1 million token context window, up from 200,000, and up to 128,000 output tokens, up from 64,000.
- **Tools:** a browser use tool on the Claude API and Google Cloud, and beta computer and browser use in the Python and TypeScript SDKs.
- **Also announced:** Sonnet 5.5's cache reads now cost $0.10 per million tokens instead of $0.20, around 20 percent cheaper on most agentic work, Anthropic says.

## How much better is Haiku 5.5 than Haiku 4.5?

On Anthropic's own benchmark table it is far ahead of Haiku 4.5 on every row where both have a figure, and still behind Sonnet 5.5 on every row.

| Benchmark (Anthropic's runs) | Haiku 4.5 | Haiku 5.5 | Sonnet 5.5 |
| --- | --- | --- | --- |
| Knowledge work, GDPval-AA v2.1 | 735 | 1620 | 1840 |
| Computer use, OSWorld 2.1 (offline subset) | 15.7 percent | 72.4 percent | 83.9 percent |
| Humanity's Last Exam, with tools | 18.7 percent | 57.4 percent | 64.5 percent |
| Agentic coding, Terminal-Bench 4.0 | 0.0 percent | 39.2 percent | 70.6 percent |
| Agentic coding, FrontierCode 1.1 (Main) | no figure | 46.4 percent | 52.1 percent (xhigh) |
| Visual reasoning, Chartography (no tools) | 6.4 percent | 46.4 percent | 61.6 percent |

Among the percentage rows, the gap to Sonnet 5.5 is smallest on FrontierCode, 46.4 against 52.1 percent, and largest on Terminal-Bench 4.0, 39.2 against 70.6. Anthropic's table also lists GPT-6 Luna, and Haiku 5.5 is ahead of it on every row where both have a figure.

## How much cheaper is Haiku 5.5, and what does a job cost on each model?

Haiku 5.5 lists at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, a 90 percent cut from Haiku 4.5's $1 and $5. Anthropic's average of around 75 percent also counts requests over 100,000 tokens, which get a smaller cut, and the newer tokenizer, which counts about 30 percent more tokens for the same text.

Here is the lineup on a made-up job: tagging 100,000 support tickets, each 400 tokens in and 50 out as Haiku 4.5 counts them, or 520 and 65 on the newer tokenizer. The prices are real list prices at standard speed, with no caching or batch discount.

| Model | Input and output, per million tokens | Cache reads, per million tokens | Cost of the made-up job |
| --- | --- | --- | --- |
| Haiku 4.5 | $1 and $5 | $0.10 | $65.00 |
| Haiku 5.5 | $0.10 and $0.50 | $0.01 | $8.45 |
| Sonnet 5.5 | $2 and $10 | $0.10 | $169.00 |
| Opus 5.5 | $4 and $20 | $0.20 | $338.00 |
| Fable 5.1 | $10 and $50 | $0.25 | $845.00 |

On this job Haiku 5.5 is 20 times cheaper than Sonnet 5.5, 40 times cheaper than Opus 5.5 and 100 times cheaper than Fable 5.1. But price is half the question. Anthropic's docs say to compare cost per completed task, because a model that fails still bills its tokens, and then the retry. On a set of science questions Haiku 5.5 cost about a twentieth of Opus 5.5 per question, not a fortieth, and scored 85 percent against 91.

## Which Claude model should you use: Fable, Opus, Sonnet or Haiku?

Start with Opus 5.5 for most workloads, as Anthropic's docs advise, reach for Haiku 5.5 when the work is high-volume and latency-sensitive, use Sonnet 5.5 for everyday coding and agent work, and keep Fable 5.1 for the hardest and longest tasks. The docs rank latency Haiku fastest, then Sonnet, Opus and Fable.

| Model | Anthropic's guidance | An example from store and marketing work (mine) |
| --- | --- | --- |
| Haiku 5.5 | High-volume, latency-sensitive work: classification, extraction, routing, summaries, subagents | Tag 100,000 support tickets, pull order numbers out of emails, route chat questions |
| Sonnet 5.5 | Speed with capability for everyday coding, data analysis, content and agentic tool use | Fix a script that pulls GA4 and store data, draft product copy, answer questions about a CSV |
| Opus 5.5 | The starting point for most workloads and the Claude Code default on most plans: long-running agentic coding, knowledge work, computer use | Rework a site's build tool, or design a tracking plan across Meta, GA4 and a store |
| Fable 5.1 | Demanding reasoning and long-horizon work, or when Opus 5.5 at higher effort falls short: hours-long sessions, deep research, ambiguous root causes. Pick it with `/model fable` | [Find out why three revenue numbers disagree](articles/reconcile-meta-ga4-and-store-revenue.html) across a quarter, or write a market report from dozens of sources |

Two rules of thumb from Anthropic. Effort is a separate dial, and the docs say tuning it is often a better lever than switching models. Its July Claude Code post adds: if Claude had the context and still got it wrong, try a more capable model, and if it skipped a file or did not run the tests, raise the effort. And compare cost per task. On a coding benchmark subset in Anthropic's own runs, Opus 5.5 at its default effort solved 92.8 percent of tasks against Fable 5.1's 92.3, for about a fifth of the cost per solved task, $0.22 against $1.19. The ranking flips by workload. The short version is also [in the FAQ](faq.html#which-claude-model-should-i-use-fable-opus-sonnet-or-haiku).

## When is Haiku 5.5 the wrong choice?

For complex agentic coding, where Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choices, and where it describes Haiku 5.5 as suited to narrower work such as compaction, summarization and subagents.

- **Complex agentic coding.** Terminal-Bench 4.0 is the sign: 39.2 percent against Sonnet 5.5's 70.6.
- **Offensive security work.** Anthropic says its cybersecurity safeguards still block penetration testing, though they allow more defensive work than Sonnet 5.5's. Requests can end in a refusal, and server-side fallback is not available for Haiku 5.5.
- **Code written for Haiku 4.5.** Manual thinking budgets, prefill and non-default sampling settings now return errors, and the same text counts as about 30 percent more tokens.

## Can one Claude model orchestrate the others?

Yes, but you assign the roles: Claude Code and the Claude Platform both let a stronger model plan or advise while cheaper models do the bulk work. In the pages I read, nothing chooses a model by task difficulty on its own; automatic switches happen only when a model is overloaded or a safety classifier flags a request.

| Option | What it does | How to start it |
| --- | --- | --- |
| Switch models | Changes the model from the next request | `/model haiku` |
| Plan, then execute | Plans with Opus and executes with Sonnet | `claude --model opusplan` |
| Subagent | Runs a delegated task on the model set in its file | `model: haiku` in the agent file |
| Advisor | The main model consults a stronger one at decision points; Anthropic API only, experimental | `/advisor opus` |
| Dynamic workflow | A script Claude writes runs many subagents, and each stage can name a model | Ask for a workflow and say which stages can use a smaller model |
| Coordinator | In Managed Agents (beta), a coordinator delegates to worker agents that each have their own model | A `multiagent` roster in the agent definition |

Anthropic's Claude Code docs list pairings. A Haiku main model with an Opus advisor is the lowest-cost main model with strong planning, costing more than Haiku alone and less than Sonnet or Opus as the main model. Cognition, the maker of Devin, says in Anthropic's launch post that Devin Fusion with Opus 5.5 as lead and Haiku 5.5 as sidekick holds a FrontierCode score of 66.2 while cutting cost and latency. That is Cognition's claim, not something I checked.

The smallest version is a subagent file, saved as `~/.claude/agents/search-term-tagger.md`:

```markdown
---
name: search-term-tagger
description: Tags a batch of ad search terms with one intent label each. Use for bulk, mechanical tagging.
tools: Read
model: haiku
---
Label each search term as brand, product, competitor, informational or other.
Return one line per term: the term, a tab, then the label. If unsure, use other.
```

The `haiku` alias resolves to Haiku 5.5 on the Anthropic API, and `/tasks` shows which model a subagent really runs on. The docs even suggest defining your own Explore subagent with `model: haiku` to explore on a cheaper model. I ran `claude plugin validate` on the file and it passes, but it also accepted a nonsense model value in my control file, so that is a format check, not a test of the routing.

Does it pay? Anthropic measured. An advisor helps only if the executor asks: in its science-question runs, Haiku 5.5 and Sonnet 5.5 executors consulted an Opus 5.5 advisor on none of 198 questions, in two runs each, gained nothing beyond noise, and paid 12 and 25 percent more for the tool's definition. An orchestrator saved money in only two measured cases, routine work with an expensive tail and work larger than one context window. Otherwise the same model at lower effort won, which is why Anthropic says to sweep effort first.

## What should you watch out for?

Four things:

- **The benchmarks are Anthropic's.** They come from its own runs and its customers' early tests, so wait for independent ones before moving something important.
- **The `haiku` alias lags on other clouds.** In Claude Code it resolves to Haiku 5.5 on the Anthropic API but to Haiku 4.5 on Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and Haiku 5.5 needs Claude Code 2.1.293 or later.
- **Orchestration is early.** The advisor is experimental and Anthropic API only, Managed Agents multi-agent is a beta, and a workflow can use meaningfully more tokens than one conversation.
- **Prices and defaults move.** Sonnet 5.5's cache price changed in this very announcement, so measure cost per completed task on your own work.

## Where this leaves me

What I like about the release is that it makes the cheap end of the range worth taking seriously: a model that scores 72.4 percent on a computer-use benchmark at a tenth of the old Haiku price changes which jobs deserve a frontier model. I would keep Opus 5.5 as the starting point, move the boring bulk work to Haiku 5.5 [behind an eval](software.html#how), and keep Fable 5.1 for the days Opus runs out of road. It is a day old and I have tested none of it, so the first job of a new model is to be wrong somewhere you can measure.

Sources: Anthropic, [Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) (October 7, 2026). Claude Platform docs, read on October 8, 2026: [What's new in Haiku 5.5](https://platform.claude.com/docs/en/models/haiku-5-5/whats-new-haiku-5-5), [Models overview](https://platform.claude.com/docs/en/models/overview), [Choosing a model](https://platform.claude.com/docs/en/about-claude/models/choosing-a-model), [Optimizing for cost and intelligence](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence), [Pricing](https://platform.claude.com/docs/en/about-claude/pricing) and [Multiagent orchestration](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration). Claude Code docs, read on October 8, 2026: [Model configuration](https://code.claude.com/docs/en/model-config), [Subagents](https://code.claude.com/docs/en/sub-agents), [Advisor tool](https://code.claude.com/docs/en/advisor) and [Dynamic workflows](https://code.claude.com/docs/en/workflows). Anthropic, [Choosing a model and effort level in Claude Code](https://claude.com/resources/articles/claude-model-and-effort-level-in-claude-code) (July 7, 2026). The models and their documentation change often, so some details here may have moved on.
