Claude Haiku 5.5 API pricing and benchmarks

What is Claude Haiku 5.5? Anthropic's model for subagents

Claude Haiku 5.5 is the newest and cheapest model in the Haiku line, Anthropic's small-model tier. It replaces Haiku 4.5 and sits below Sonnet 5.5 and Opus 5.5 in the Claude 5.5 generation.

Anthropic pitches it at work that runs in bulk and where cost matters more than depth. It is meant to work alongside the larger models rather than replace them.

Anthropic API model ID and cloud availability

On the Claude Platform, you call it with the model ID claude-haiku-5-5. It went live there on October 7, and Amazon Web Services, Google Cloud, and Microsoft Azure added it the same day.

Anthropic has published a migration guide for teams moving off Haiku 4.5. Read it before you swap the model ID in production code.

Subagents and other high-volume jobs Anthropic recommends

Anthropic recommends Haiku 5.5 as a subagent that takes orders from Opus 5.5 or Sonnet 5.5 during coding work. A subagent is a helper model that handles the small, well-defined pieces of a plan a larger model has drawn up.

Anthropic lists other jobs as well. All of them are quick, repetitive, and run at high volume:

• Summarizing documents and conversations
• Compaction, which shrinks a long conversation history
• Database queries
• Classifying incoming requests

Because it is Anthropic's quickest model at standard speed, the company also points it at live customer support and browser automation. Anthropic describes both as speed-sensitive.

Claude Haiku 5.5 API pricing vs Haiku 4.5

Input price per million tokens (USD)

Bar heights are scaled to Sonnet 5.5's $2.00.

0.10Haiku 5.5 (≤100K)

0.50Haiku 5.5 (>100K)

1.00Haiku 4.5

2.00Sonnet 5.5

00.511.52

Haiku 5.5 has two price tiers, and which one you pay depends on whether a prompt stays at or under 100K tokens. API usage is billed in tokens, the small text chunks a model reads and writes. Prices below are per million tokens, split into input you send and output you get back.

Under 100K tokens, API pricing is 90% lower than Haiku 4.5

For prompts of 100K tokens or fewer, both input and output cost one tenth of what Haiku 4.5 charges. Anthropic says roughly 90% of Haiku 4.5 traffic fell in that range. Cache reads are the cheaper rate you pay when a prompt reuses content sent earlier.

Here is what that means in practice. Run 10,000 calls of 2,000 input tokens and 500 output tokens each, and assume the token counts do not change: the job costs $45 on Haiku 4.5 and $4.50 on Haiku 5.5.

Over 100K tokens, pricing is still half of Haiku 4.5's

Once a prompt passes 100K tokens, the rate climbs to $0.50 input and $2.50 output. That is five times the lower tier and only half of Haiku 4.5's price. Input at that rate is still a quarter of what Sonnet 5.5 charges.

Token counts also shift. Haiku 5.5 ships with a new tokenizer, the component that splits text into tokens, and it produces slightly more tokens for the same job. Anthropic's 75% average already folds in both that increase and the share of prompts above 100K.

Haiku 5.5 test scores against three rival models

Benchmarks published by Anthropic

OSWorld 2.1 (%)

Operating a real computer through its screen (offline subset).

72.4Haiku 5.5

15.7Haiku 4.5

48.9GPT-6 Luna

83.9Sonnet 5.5

0255075

Terminal-Bench 4.0 (%)

Multi-step coding tasks in a command-line interface.

39.2Haiku 5.5

0.0Haiku 4.5

16.4GPT-6 Luna

70.6Sonnet 5.5

0204060

GDPval-AA v2.1 (Elo)

Professional work across 44 occupations. Bars scaled so Sonnet 5.5 = 100.

1620Haiku 5.5

735Haiku 4.5

1437GPT-6 Luna

1840Sonnet 5.5

05001,0001,500

Haiku 5.5 outscores both Haiku 4.5 and OpenAI's GPT-6 Luna on OSWorld 2.1 and Terminal-Bench 4.0, but it stays well behind Sonnet 5.5 on coding. All figures come from Anthropic's launch post, which charts Haiku 5.5 against those three models.

Benchmark gains over Haiku 4.5

Wherever Anthropic lists a Haiku 4.5 score, Haiku 5.5 beats it. The biggest jump is on OSWorld 2.1, a test of driving a computer through its screen, where the score rises from 15.7% to 72.4%.

Humanity's Last Exam, a set of expert-level academic questions, goes from 10.2% to 45.9% without tools and from 18.7% to 57.4% with them. Chartography, which checks how well a model reads charts, moves from 6.4% to 46.4%.

Where Sonnet 5.5 and GPT-6 Luna land

Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, a command-line coding test, against Haiku 5.5's 39.2%. Sonnet also leads on GDPval-AA v2.1, which rates professional work across 44 occupations on an Elo scale, at 1840 to 1620.

Against GPT-6 Luna, OpenAI's small model, Haiku 5.5 comes out ahead on both computer use (72.4% vs 48.9% on OSWorld 2.1) and Terminal-Bench 4.0 (39.2% vs 16.4%). Luna is the only non-Anthropic model in the launch charts.

Haiku vs Sonnet 5.5 for coding and subagents

For complex agentic coding, where the model plans and carries out multi-step coding work on its own, Anthropic still points developers to Sonnet 5.5 or Opus 5.5. Haiku 5.5 is the model for tightly scoped jobs that earlier Claude versions made too expensive to run at volume. Summaries, compaction, and subagent tasks are the examples Anthropic gives.

Sonnet 5.5 cache read pricing cut in half

On the same day, Anthropic halved what Sonnet 5.5 charges for cache reads, from $0.20 to $0.10 per million tokens. Agent workloads re-read cached context constantly, so those reads make up a large share of the tokens they use.

Anthropic estimates the cut makes Sonnet 5.5 about 20% cheaper on most agentic tasks. Our Claude Sonnet 5.5 vs Opus 5.5: price, benchmarks article covers the rest of that model.

Haiku 5.5 subagents under Opus 5.5 or Sonnet 5.5

The setup Anthropic has in mind uses both tiers: Opus 5.5 or Sonnet 5.5 plans and decides, and Haiku 5.5 subagents handle the lookups and summaries. You do not have to choose one model for the whole pipeline.

Rogo, one of the early customers quoted at launch, describes exactly this split. A larger model assembles a slide deck while a Haiku 5.5 subagent pulls the segment revenue figure the deck needs from a company's 10-K filing. For the top tier, see Claude Opus 5.5 pricing, benchmarks, API changes.

Claude Haiku 5.5 API settings: model ID, effort, and credits

Point your API calls at claude-haiku-5-5, then pick an effort level that matches the task. If you moved over from Haiku 4.5, compare your code against Anthropic's migration guide before you ship.

Anthropic also updated its Python and TypeScript SDKs on launch day, adding computer use and browser use in beta. It names Haiku 5.5 as a strong fit for those tasks because of its speed and price.

Effort settings and API pricing per task

Haiku 5.5 is the first Haiku model with an adjustable effort setting, offering five levels from Low to Max. Effort controls how much the model reasons before it answers.

Lower effort means less reasoning and fewer tokens, so each task costs less. Higher effort spends more tokens for a better answer, and you can set it per task.

Anthropic's monthly API credits for Max and Team

Starting this week, Claude Max and Team subscribers get a monthly credit for the Claude Platform API. The amounts depend on the plan:

• Max 5x: $100 a month
• Max 20x: $200 a month
• Team: up to $500, shared across the team's users

The credits are not tied to Haiku 5.5 and work with any Claude model. Anthropic intends them for trying out tools, apps, and agents built on the API.

Haiku 5.5 costs far less on prompts under 100K tokens

Claude Haiku 5.5 charges a tenth of Haiku 4.5's per-token price on prompts up to 100K tokens and scores far higher on Anthropic's benchmarks. For summaries, classification, and subagent work, switching cuts the bill sharply.

The savings shrink on prompts above 100K tokens, where prices rise fivefold, and Sonnet 5.5 remains the stronger model for complex coding. Check how many tokens your typical request sends before you migrate.

Our model comparison lines up all four Claude tiers side by side, with prices for each. Read Claude models compared: Fable, Opus, Sonnet, Haiku to see where Haiku 5.5 fits.