Claude Haiku 5.5 API

Claude Haiku 5.5 focuses on high-volume classification, extraction, routing and subagent work. It supports adaptive reasoning, but its token prices change when the complete input exceeds 100,000 tokens.

MODALITIES

TextText

Available

INPUT / OUTPUT70% OFF

$0.03/ 1M

$0.15/ 1M

USD per million tokens

CACHED INPUT

$0.003/ 1M

Platform price · USD

CONTEXT WINDOW

1,000,000tokens

Max output: 128,000 tokens

Claude Haiku 5.5

API Usage Get API access

API Usage Examples

API documentation ↗

POSTGet API Key ↗

curl 'https://newrouters.com/v1/messages' \
  -H "Content-Type: application/json" \
  -H "x-api-key: $MODEL_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  --data '{
  "model": "claude-haiku-5-5",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "Explain how an HTTP request works in three concise steps."
    }
  ],
  "stream": false
}'

const apiKey = process.env.MODEL_API_KEY;

const response = await fetch("https://newrouters.com/v1/messages", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    "x-api-key": apiKey,
    "anthropic-version": "2023-06-01"
  },
  body: JSON.stringify({
    "model": "claude-haiku-5-5",
    "max_tokens": 1024,
    "messages": [
      {
        "role": "user",
        "content": "Explain how an HTTP request works in three concise steps."
      }
    ],
    "stream": false
  })
});

if (!response.ok) throw new Error(await response.text());
console.log(await response.text());

import os
import json
import requests

response = requests.post(
    "https://newrouters.com/v1/messages",
    headers={
        "Content-Type": "application/json",
        "x-api-key": os.environ["MODEL_API_KEY"],
        "anthropic-version": "2023-06-01"
    },
    json=json.loads("{\"model\":\"claude-haiku-5-5\",\"max_tokens\":1024,\"messages\":[{\"role\":\"user\",\"content\":\"Explain how an HTTP request works in three concise steps.\"}],\"stream\":false}"),
    timeout=300,
)
response.raise_for_status()
print(response.text)

Set MODEL_API_KEY in your server environment. Requests use this model’s native API and your account pricing.

Docs: authentication and keys · Error handling

Claude Haiku 5.5 pricing

Estimate one request, then choose a budget for your workflow.

Usage	Quantity	Rate	Cost
Input · input ≤ 100,000	tokens	$0.03/ 1M	$0.00003
Cached input · input ≤ 100,000	tokens	$0.003/ 1M	$0
Cache write · 5m · input ≤ 100,000	tokens	$0.0375/ 1M	$0
Cache write · 1h · input ≤ 100,000	tokens	$0.06/ 1M	$0
Output · input ≤ 100,000	tokens	$0.15/ 1M	$0.000075
Reasoning · input ≤ 100,000	tokens	$0.15/ 1M	$0

Cost per request$0.000105

Public USD estimate. Your account rate and final recorded usage determine the actual charge.

Model pricing details

Specification	Our price	Official reference
Input · input ≤ 100,000	$0.03 / 1M tokens	$0.1 About 70% less than official
Input · input > 100,000	$0.15 / 1M tokens	$0.5 About 70% less than official
Cached input · input ≤ 100,000	$0.003 / 1M tokens	$0.01 About 70% less than official
Cached input · input > 100,000	$0.015 / 1M tokens	$0.05 About 70% less than official
Cache write · 5m · input ≤ 100,000	$0.0375 / 1M tokens	$0.125 About 70% less than official
Cache write · 5m · input > 100,000	$0.1875 / 1M tokens	$0.625 About 70% less than official
Cache write · 1h · input ≤ 100,000	$0.06 / 1M tokens	$0.2 About 70% less than official
Cache write · 1h · input > 100,000	$0.3 / 1M tokens	$1 About 70% less than official
Output · input ≤ 100,000	$0.15 / 1M tokens	$0.5 About 70% less than official
Output · input > 100,000	$0.75 / 1M tokens	$2.5 About 70% less than official
Reasoning · input ≤ 100,000	$0.15 / 1M tokens	$0.5 About 70% less than official
Reasoning · input > 100,000	$0.75 / 1M tokens	$2.5 About 70% less than official

Related Models

All Models ↗

Model	Input / output	Cached input	Context	Input → output
Claude Haiku 5.5CurrentAnthropic	$0.03 / $0.15per 1M tokens	$0.003per 1M tokens	1,000,000	TextText
Claude Fable 5Same claude protocolSame billing unitsAnthropic	$3 / $15per 1M tokens	$0.3per 1M tokens	1,000,000	TextText
Claude Fable 5.1Same claude protocolSame billing unitsAnthropic	$3 / $15per 1M tokens	$0.075per 1M tokens	1,000,000	TextText
Claude Opus 4.6Same claude protocolSame billing unitsAnthropic	$1.5 / $7.5per 1M tokens	$0.15per 1M tokens	1,000,000	TextText
Claude Opus 4.7Same claude protocolSame billing unitsAnthropic	$1.5 / $7.5per 1M tokens	$0.15per 1M tokens	1,000,000	TextText
Claude Opus 4.8Same claude protocolSame billing unitsAnthropic	$1.5 / $7.5per 1M tokens	$0.15per 1M tokens	1,000,000	TextText

Claude Haiku 5.5 API · Examples

Original prompts adapted to documented capabilities; these are not vendor-generated outputs.

Official source ↗ · 2026-10-15

EXAMPLE PROMPT

Analyze this material. Separate supported conclusions, conflicting evidence and unanswered questions. Cite the relevant passages: [material]

USER MESSAGE

Evidence-based analysis

Verify claims against the supplied evidence.

EXAMPLE PROMPT

Extract entities and events from this text into a JSON structure. Use null for missing facts and do not infer identities: [text]

USER MESSAGE

Structured extraction

Constrain the output and handle missing information explicitly.

EXAMPLE PROMPT

Plan a tool-assisted workflow for this task. State the inputs for each call, when to ask for clarification, and how to verify completion: [task]

USER MESSAGE

Tool workflow design

Make the workflow and stopping conditions explicit.

EXAMPLE PROMPT

Review this function for empty inputs, duplicate calls and concurrency races. Propose a minimal fix and focused verification: [code]

USER MESSAGE

Boundary-case review

Assess a concrete implementation with checkable cases.

What Is Claude Haiku 5.5?

Official model information ↗ · 2026-10-15

Official model specifications

Official context window	1,000,000 tokens
Synchronous output limit	128,000 tokens
Input / output	Text and images → text
Knowledge cutoff	June 2026
Minimum cacheable prompt	512 tokens
Default effort	medium
Official release	2026-10-07

Choose this model

Use it for classification, routing, extracting structured fields and focused subagent tasks. Keep the task and output schema precise; long prompts may change the applicable price tier.

Thinking and migration

Adaptive thinking is enabled by default with medium effort. Thinking blocks are account-bound unless the accounts are linked. Keep multi-turn tool interactions on a compatible account and preserve returned content blocks.

Context, output and modalities

Official specifications list a 1M-token context window and 128k synchronous output. Text and image input produce text output. The 300k output option uses the vendor’s Message Batches beta and is a separate API capability; it is not the synchronous output limit advertised here.

Prompt caching

The minimum cacheable prompt is 512 tokens. Cache writes have separate 5-minute and 1-hour rates; subsequent reads use the cache-read rate. Cache eligibility does not guarantee a hit: prompt prefixes, expiration and upstream support still matter.

Native API integration

Use the public model ID with POST /v1/messages and x-api-key authentication. Streaming uses the native Claude event format. Preserve tool_use and tool_result messages for multi-turn tool calls. Query the platform contract before relying on vendor-managed tools or optional beta features.

Sampling and operational limits

Do not reuse sampling settings from older models without checking migration guidance. Prefer omitting temperature, top_p and top_k. Thinking consumes output tokens, so leave enough output budget for both reasoning and the final answer. Official capabilities are distinct from current platform availability.

Official pricing and platform discount

Standard input / output rates are USD 0.10 / 0.50; above 100k input: 0.50 / 2.50 per million tokens. The 5-minute write / 1-hour write / cache-read rates are USD 0.125 / 0.20 / 0.01; above 100k: 0.625 / 1 / 0.05 per million tokens. The live platform table is authoritative for the published 0.3 multiplier. Batch, Fast mode, regional inference and managed tools can have separate vendor charges.

The 100k input boundary

The threshold counts uncached input, cache-read input and cache-write input together. At exactly 100,000 input tokens the lower tier applies; above it the higher rates apply to the whole request, including output. The platform selects the applicable tier from final recorded input usage and applies the published multiplier to that tier.

Official references

Reviewed on 2026-10-15

Claude Haiku 5.5 Platform specifications

Specification	Claude Haiku 5.5
Model ID	claude-haiku-5-5
Provider	Anthropic
API	/v1/messages
Context window	1,000,000 tokens
Maximum output	128,000 tokens
Declared capabilities	Tools · Streaming · Reasoning · Structured output

Protocol & capability details

Protocol & capabilities

claude · claude-haiku-5-5

Context limit: 1000000
Maximum output: 128000

tools · stream · reasoning · structured_output

Compare Claude Haiku 5.5 with related models

Specification	Claude Haiku 5.5	Claude Fable 5	Claude Fable 5.1
Provider	Anthropic	Anthropic	Anthropic
Input → output	TextText	TextText	TextText
Selection basis	Current model	Same claude protocol · Same billing units	Same claude protocol · Same billing units
Context / max output	1,000,000 / 128,000	1,000,000 / 128,000	1,000,000 / 128,000
Starting price	$0.03 / 1M tokens	$3 / 1M tokens	$3 / 1M tokens

Choose by the input types, output settings and billing unit your workflow needs. Try the same prompt across models to compare results for your own use case.

Integrate Claude Haiku 5.5 into your app

Keep the native request format and point your application to the platform API. Prepare a key with model access, then review status and charges in your usage records.

View API examples

Claude Haiku 5.5 API · Common Questions

Everything you need to know about API access, pricing and model capabilities.

What is the default reasoning effort?

The default is medium. Adaptive thinking settings differ from older Claude models.

Is the vendor’s batch output limit available here?

The vendor’s 300k batch beta is distinct from the synchronous Messages endpoint. Only platform-documented endpoints and limits are available here.

Does the starting price apply to 1M-token prompts?

No. Prompts above 100k input tokens use the higher official tier for input, caching and output. The platform applies the higher tier when the recorded input crosses that threshold.

How is Claude Haiku 5.5 billed?

Charges depend on the recorded token usage and your account rates. Input, output and supported cache or reasoning categories are listed separately in the pricing table. The calculator uses public USD prices; the final usage receipt determines the charge.

How do I get an API key?

Choose “Get API access” to open the integration guide. Sign in to your NewRouters account, prepare a key with access to this model and copy an API example.

Which API endpoints can I use?

Supported endpoints: /v1/messages、/v1/messages/count_tokens. Choose an endpoint in the API examples to see its native request format.

Does the model support streaming?

The current model route declares streaming support. Use the native endpoint’s streaming option and consume its events. Usage already completed may still be billed if the connection is interrupted.

Can I use tools or structured output?

Tools: supported. Structured output: supported. Request fields follow the selected native endpoint.

Where can I check usage and charges?

Keep the request ID and open your account’s LLM usage records to inspect request status, token counts and charges. Prompt and answer content are not a conversation history in this view.

Explore LLM Use Cases

All LLM models ↗

Integration topics / Documentation

All documentation ↗

QUICK STARTHello, API.

Docs: Make your first API request

Prepare a key, choose a model and send a minimal request.

View docs section ↗

NATIVE API{ messages: [...] }

Docs: Use the native API

Explore request formats for OpenAI, Claude and Gemini.

View docs section ↗

INTEGRATIONbuild(something)

Docs: Choose your first request endpoint

Configure requests, handle responses and review usage.

View docs section ↗