New Model Added: Claude Haiku 5.5

Date: 2026-10-13 / (1405-07-21)

Summary

Anthropic's Claude Haiku 5.5 is now available on AvalAI for high-volume, cost-sensitive tasks and coding subagents. It supports Chat Completions, Messages, and Responses, with adjustable reasoning effort and lower token rates for requests up to 100K input tokens.

Details

Anthropic: Claude Haiku 5.5

Use claude-haiku-5-5 for summarization, conversation compaction, database queries, classification, quick lookups, and live customer support. It can handle narrowly scoped subagent work while Sonnet 5.5 or Opus 5.5 coordinates the larger task. Anthropic provider guide

• High-volume work: designed for short, repetitive tasks where response speed and cost matter.
• Adjustable effort: Anthropic's first Haiku-class model with an effort setting to balance cost and intelligence. Evaluate effort on your own tasks rather than inheriting another Claude model's default.
• Computer and browser use: Anthropic reports improved computer-use performance. Model capability does not automatically enable provider-hosted browser tools or beta SDK features on an AvalAI route.
• Provider-reported speed and cost: Anthropic calls it its fastest model at standard speed and reports around 75% lower average task cost than Haiku 4.5. Its Opus Fast Mode remains faster. These are upstream, workload-dependent comparisons, not AvalAI latency or cost-per-task guarantees.
• Updated tokenizer: token counts can be slightly higher than Haiku 4.5 for the same work. Re-measure prompt size, output budgets, and cost before migrating.

Selected provider benchmarks

The following results are reported by Anthropic, not independently measured on AvalAI. Keep evaluation versions and tool conditions in mind; they do not predict every production workload.

Evaluation	Haiku 5.5	Haiku 4.5
GDPval-AA v2.1, knowledge work	1620	735
OSWorld 2.1, offline subset	72.4%	15.7%
Humanity's Last Exam, no tools	45.9%	10.2%
Humanity's Last Exam, with tools	57.4%	18.7%
Terminal-Bench 4.0, agentic coding	39.2%	0.0%

Sonnet 5.5 and Opus 5.5 remain better choices for complex, open-ended agentic coding. See the Haiku 5.5 system card for evaluation methodology and safety findings.

Supported endpoints

Model ID	Endpoint
claude-haiku-5-5	v1/chat/completions
claude-haiku-5-5	v1/messages
claude-haiku-5-5	v1/responses

Use OpenAI-compatible clients for Chat Completions and Responses, or the native Anthropic SDK for Messages. Endpoint availability does not establish support for every hosted tool, stored conversation feature, Batch option, or provider beta. This release does not automatically replace existing model IDs or aliases.

AvalAI pricing

All prices below are USD per 1 million tokens for claude-haiku-5-5.

Total request input length	Input	Cached input	Cache creation input	Output
Up to and including 100K tokens	$0.10	$0.01	$0.20	$0.50
Above 100K tokens	$0.50	$0.05	$1.00	$2.50

Exactly 100K input tokens stays in the lower tier. Total request input length selects the applicable rates, including output; the higher tier is not a marginal surcharge on only the tokens beyond 100K. 100K is a pricing threshold, not a context-window limit. Confirm current context, output, and account-tier limits in the live model catalog.

Use AvalAI's cache-creation rates of $0.20 / $1.00, not the upstream $0.125 / $0.625 cache-write prices. These rates do not establish a cache lifetime. Pricing details

Anthropic's accompanying Sonnet 5.5 cache-read discount and Claude Max/Team subscription credits concern the Claude Platform. They are not announced AvalAI pricing changes or AvalAI account credits here.

Chat Completions request and SDK examples

Set AVALAI_API_KEY in your server environment. Start with a minimal request before adding route-specific reasoning or tool settings.

Bash

curl https://api.avalai.ir/v1/chat/completions \
  -H "Authorization: Bearer $AVALAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-haiku-5-5",
    "messages": [{"role": "user", "content": "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout."}]
  }'

Python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AVALAI_API_KEY"], base_url="https://api.avalai.ir/v1"
)
response = client.chat.completions.create(
    model="claude-haiku-5-5",
    messages=[
        {
            "role": "user",
            "content": "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout.",
        }
    ],
)
print(response.choices[0].message.content)

JavaScript

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.AVALAI_API_KEY,
  baseURL: "https://api.avalai.ir/v1",
});
const response = await client.chat.completions.create({
  model: "claude-haiku-5-5",
  messages: [{ role: "user", content: "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout." }],
});
console.log(response.choices[0].message.content);

Illustrative response

This is an illustrative response, not a captured live API result. It assumes 100 uncached prompt tokens, 50 billed completion tokens, no tools, and an example exchange rate of 100,000 IRT/USD. Actual identifiers, token counts, reasoning details, and exchange rates vary. At the lower pricing tier, the example costs (100 × $0.10 + 50 × $0.50) / 1,000,000 = $0.000035, or 3.5 IRT at that example rate.

{
  "id": "chatcmpl-claude-haiku-5-5-example",
  "created": 1791849600,
  "model": "claude-haiku-5-5",
  "object": "chat.completion",
  "system_fingerprint": null,
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "The customer reports a duplicate charge after retrying checkout.",
        "thinking_blocks": [],
        "annotations": []
      }
    }
  ],
  "usage": {
    "prompt_tokens": 100,
    "completion_tokens": 50,
    "total_tokens": 150,
    "completion_tokens_details": null,
    "prompt_tokens_details": {
      "cached_tokens": 0,
      "text_tokens": 100,
      "audio_tokens": null,
      "image_tokens": null
    }
  },
  "estimated_cost": {
    "unit": "0.0000350000",
    "irt": 3.5,
    "exchange_rate": 100000
  }
}

Messages and adjustable effort

Use a host-only base URL for the Anthropic SDK, without /v1. This example selects low through the native output_config.effort control for a short task. Verify accepted effort values on your route before using an override; omit output_config for a baseline request. Do not assume Haiku inherits Sonnet's or Opus's default effort, thinking-off behavior, or fixed thinking budgets.

Bash

curl https://api.avalai.ir/v1/messages \
  -H "x-api-key: $AVALAI_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-haiku-5-5",
    "max_tokens": 4096,
    "output_config": {"effort": "low"},
    "messages": [{"role": "user", "content": "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout."}]
  }'

Python

import os
from anthropic import Anthropic

client = Anthropic(
    api_key=os.environ["AVALAI_API_KEY"], base_url="https://api.avalai.ir"
)
response = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=4096,
    output_config={"effort": "low"},
    messages=[
        {
            "role": "user",
            "content": "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout.",
        }
    ],
)
for block in response.content:
    if block.type == "text":
        print(block.text)

JavaScript

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  apiKey: process.env.AVALAI_API_KEY,
  baseURL: "https://api.avalai.ir",
});
const response = await client.messages.create({
  model: "claude-haiku-5-5",
  max_tokens: 4096,
  output_config: { effort: "low" },
  messages: [{ role: "user", content: "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout." }],
});
for (const block of response.content) {
  if (block.type === "text") console.log(block.text);
}

Reasoning consumes billed output tokens and the output budget. Choose the lowest supported effort that passes your evaluations, increase it only when quality improves, and retain enough room for the final answer. Preserve complete assistant and tool turns for subsequent requests; filtering text for display is not a safe history-storage format. Ask for concise conclusions or verification, not hidden chain-of-thought.

Responses example

This is a minimal stateless request. Test any hosted-tool or stateful workflow separately rather than inferring it from endpoint availability.

Bash

curl https://api.avalai.ir/v1/responses \
  -H "Authorization: Bearer $AVALAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-haiku-5-5",
    "input": "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout."
  }'

Python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["AVALAI_API_KEY"], base_url="https://api.avalai.ir/v1"
)
response = client.responses.create(
    model="claude-haiku-5-5",
    input="Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout.",
)
print(response.output_text)

JavaScript

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.AVALAI_API_KEY,
  baseURL: "https://api.avalai.ir/v1",
});
const response = await client.responses.create({
  model: "claude-haiku-5-5",
  input: "Summarize this support ticket in one sentence: the customer was charged twice after retrying checkout.",
});
console.log(response.output_text);

Migration checklist

1. Select the explicit claude-haiku-5-5 ID; do not assume an existing Haiku alias now points to it.
2. Follow the official migration guide and verify request controls and model limits before switching traffic.
3. Re-test classification accuracy, summaries, tool arguments, latency, and tokenizer-dependent usage on representative tasks.
4. Apply the 100K pricing threshold to the entire request estimate, including output and cache operations.
5. Keep database queries and browser actions behind validation, permissions, and approval boundaries. Delegate narrowly scoped work to Haiku; keep difficult planning with Sonnet 5.5 or Opus 5.5.

Related documentation

• Anthropic models
• Chat Completions, Messages, and Responses
• Reasoning guide
• AvalAI pricing and prompt caching
• Haiku 5.5 system card