What are the strengths of this model?

Matches GPT-6 Luna on price while beating it on agentic benchmarks First Haiku with an effort dial and adaptive thinking on by default 1M context and vision input at $0.10/$0.50 per 1M Halves Sonnet 5.5 cache-read price to $0.10, helping agent loops

What are the weaknesses of this model?

$0.10/$0.50 only holds below 100K tokens (5x above) New tokenizer counts ~30% more, so real savings near 75% not 90% Requires omitting temperature/top_p/top_k; existing Haiku 4.5 code needs retesting Trails Sonnet/Opus on complex coding

What are the best use cases?

High-volume sub-agent execution, classification and routing Live support and in-app assistants that read images Browser and desktop automation (OSWorld 72.4%) Triage in front of a stronger model

AnthropicProprietary

Claude Haiku 5.5

Name: Claude Haiku 5.5
Author: Anthropic

Compare this model

Anthropic's Claude Haiku 5.5, released Oct 7, 2026, is the cheapest model in the Claude 5 family and the first Haiku with reasoning. It carries a 1M-token context window with vision input and is priced at the same $0.10/$0.50 per 1M tokens as OpenAI's GPT-6 Luna, launched the same day. With an effort dial and adaptive thinking on by default, it beats Luna on agentic work — 72.4% on OSWorld and 39.2% on Terminal-Bench 4.0.

Share:X Hatena LINE

Parameters

undisclosed (Haiku class)

Context Window

License

Proprietary

Release Date

2026-10-07

Benchmark Performance

AA Intelligence Index

34.0

LMArena Elo

—

HLE

45.9

ARC-AGI-2

—

SWE-bench Verified

—

GPQA Diamond

—

MMLU-Pro

—

LiveCodeBench

—

AIME 2025

—

MATH-500

—

API Pricing

Input Price (per 1M tokens)

$0.1

Output Price (per 1M tokens)

$0.5

Billing Mode: per 1M tokens (prompts <=100K; 5x above)

Strengths

•Matches GPT-6 Luna on price while beating it on agentic benchmarks
•First Haiku with an effort dial and adaptive thinking on by default
•1M context and vision input at $0.10/$0.50 per 1M
•Halves Sonnet 5.5 cache-read price to $0.10, helping agent loops

Weaknesses

•$0.10/$0.50 only holds below 100K tokens (5x above)
•New tokenizer counts ~30% more, so real savings near 75% not 90%
•Requires omitting temperature/top_p/top_k; existing Haiku 4.5 code needs retesting
•Trails Sonnet/Opus on complex coding

Use Cases

•High-volume sub-agent execution, classification and routing
•Live support and in-app assistants that read images
•Browser and desktop automation (OSWorld 72.4%)
•Triage in front of a stronger model

Deep Analysis

Artificial Analysis Intelligence Index

43 (max) / 34 (default)

measured at max and medium effort; vs 17 for Haiku 4.5 with reasoning

Terminal-Bench 4.0

39.2%

vs 0.0% Haiku 4.5 and 16.4% GPT-6 Luna; Sonnet 5.5 leads at 70.6%

OSWorld 2.1 (offline subset)

72.4%

desktop control; vs 15.7% Haiku 4.5 and 48.9% GPT-6 Luna

Input / Output Price

$0.10 / $0.50 per 1M

prompts up to 100K tokens; 5x above 100K ($0.50/$2.50); cache read $0.01

Context Window

1M tokens

128K max output (300K via Batch beta); vision input; cutoff June 2026

HLE (no tools / with tools)

45.9% / 57.4%

expert-level exam, self-reported at max effort

Strengths

・First Haiku with an adjustable effort dial and adaptive thinking on by default, so it can actually reason instead of only classify
・Matches GPT-6 Luna price to the cent while beating it on agentic work: 72.4% vs 48.9% on OSWorld, 39.2% vs 16.4% on Terminal-Bench
・1M context and vision input at $0.10/$0.50 make it the cheapest credible multimodal model with a real long-context window
・Cutting Sonnet 5.5 cache-read pricing to $0.10 (from $0.20) saves agent-loop builders more than the Haiku headline discount does

Weaknesses

・The $0.10/$0.50 rate only holds below 100K tokens; above that input jumps 5x and the new tokenizer counts roughly 30% more tokens, so real savings on long prompts land near 75% not the advertised 90%
・Modal constraint: temperature, top_p and top_k must be omitted and assistant prefill is rejected with a 400; Priority Tier is unsupported, so Haiku 4.5 integrations need retesting
・It is still a Haiku: Sonnet 5.5 and Opus 5.5 remain the recommended models for complex coding agents, and Haiku 5.5 trails them on reasoning composites

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GPT-6 Luna	34 (AA Intelligence, default)	16.4% (Terminal-Bench 4.0)	N/A	$0.10 / $0.50 per 1M
Claude Sonnet 5.5	64 (AA Intelligence)	70.6% (Terminal-Bench 4.0)	N/A	$3 / $15 per 1M
Claude Haiku 4.5	17 (AA Intelligence, reasoning)	0.0% (Terminal-Bench)	N/A	$1 / $5 per 1M

Overview

Claude Haiku 5.5 is Anthropic's smallest, cheapest, fastest model and the first Haiku that can reason. Released on October 7, 2026 at exactly the same $0.10 / $0.50 per 1M tokens as OpenAI's GPT-6 Luna, it is built for the high-volume jobs you run thousands of times a day: extraction, classification, routing, subagents, live support. The model ID is claude-haiku-5-5; it reads text and images, holds a 1M-token context window, writes up to 128K tokens per answer, and is the first Haiku with an effort dial (low / medium / high / xhigh / max, default medium) plus adaptive thinking on by default. Anthropic's own table shows Haiku 5.5 beating Luna on every shared benchmark, most sharply on computer and terminal work: 72.4% vs 48.9% on OSWorld 2.1 and 39.2% vs 16.4% on Terminal-Bench 4.0 (Haiku 4.5 scored 0.0% on both). Artificial Analysis places it at 43 on its Intelligence Index at max effort and 34 at the default medium setting, against 17 for Haiku 4.5 with reasoning. The honest caveats are the price tier and the tokenizer: the headline rate only holds below 100K tokens, and the new tokenizer counts about 30% more tokens than Haiku 4.5, so Anthropic itself settles the real saving at roughly 75% rather than the advertised 90%.

Benchmarks & Performance

Detailed Comparison

Community Feedback

Reaction on launch day was unusually sharp because of the Luna price match. The recurring observation across developer forums was that two labs announcing a day apart and landing on the same number to the cent is not competition, it is a floor. * **What people liked.** The adaptive-thinking arrival on a Haiku class model drew the most comment. Reviewers noted that for the first time a Haiku can do real agentic steps instead of only classification, and the Terminal-Bench jump from 0.0% to 39.2% is a genuine capability change, not a marketing line. * **What people pushed back on.** The tokenizer math. Several teams recalculated that the 90% cheaper claim settles near 75% once the new tokenizer and the 100K tier are accounted for, and that roughly 90% of Haiku 4.5 traffic stayed under 100K tokens, so most users do land in the cheap bracket but not at the advertised discount. * **Migration friction.** Developers warned that Haiku 5.5 rejects temperature, top_p and top_k, rejects assistant prefill with a 400, and does not support Priority Tier. Existing Haiku 4.5 code that set those parameters breaks and needs retesting. * **The Sonnet discount.** The quiet masterstroke many missed: cutting Sonnet 5.5 cache reads to $0.10 saves agent-loop builders more than the Haiku headline number does, because cache reads dominate long agent runs.

Use Cases

**1. High-volume subagent execution** * Example: an orchestrator farms out thousands of extraction, lookup and classification steps per hour to Haiku 5.5 subagents. * Why Haiku 5.5: $0.10/$0.50 at default effort and a 1M window make it the cheapest frontier-lab worker that can also follow tool calls; keep each prompt under 100K tokens to stay in the cheap tier. **2. Live support and in-app assistants** * Example: a customer-support bot that classifies intent, pulls a record and replies in under a second. * Why Haiku 5.5: vision input plus adaptive thinking let it read a screenshot or a form and reason briefly; 128K output covers long structured replies. **3. Browser and desktop automation** * Example: an agent that drives a settings page or a checkout flow. * Why Haiku 5.5: 72.4% on OSWorld and 39.2% on Terminal-Bench are the best a Haiku has ever posted, good enough for narrow, repeatable UI tasks. **4. Routing and triage in front of a stronger model** * Example: a classifier decides which queries escalate to Sonnet 5.5 or Opus 5.5. * Why Haiku 5.5: cheap enough to sit in front of every request; the effort dial lets it spend a little more only when a task looks hard. **When to pick something else** * Sonnet 5.5 or Opus 5.5 for complex coding agents and multi-step debugging. * An open-weight small model if you need on-prem control and the 100K token price floor does not matter to you.

Latest News

**Launch (October 7, 2026).** Anthropic released Claude Haiku 5.5 as the cheapest model in the Claude 5 family and the first Haiku with reasoning. It went live the same day on the Claude API (model ID claude-haiku-5-5), Claude.ai on all plans, Claude Code, Amazon Bedrock, Claude Platform on AWS, Google Cloud Vertex AI and Microsoft Foundry. Anthropic committed to keeping it available until at least October 7, 2027. **Price held to the cent against Luna.** $0.10 per million input tokens and $0.50 per million output for prompts up to 100K tokens; $0.50 / $2.50 above that; cache read $0.01 ($0.05 above 100K); Batch API takes 50% off. The same-day GPT-6 Luna launch used the identical $0.10 / $0.50 list price. **First Haiku with an effort dial.** Five levels (low / medium / high / xhigh / max), default medium, with adaptive thinking on by default. Knowledge cutoff is June 2026; modalities are text and image in, text out. **Breaking changes that break existing code.** temperature, top_p and top_k must be omitted (non-default values return 400); assistant prefill is rejected with a 400 even with thinking off; Priority Tier is not supported; sending thinking budget_tokens returns a 400. Migrations from Haiku 4.5 need retesting.

Anthropic's own table shows Haiku 5.5 beating Luna on every shared benchmark, most sharply on computer and terminal work: 72.4% vs 48.9% on OSWorld 2.1 and 39.2% vs 16.4% on Terminal-Bench 4.0 (Haiku 4.5 scored 0.0% on both). Artificial Analysis places it at 43 on its Intelligence Index at max effort and 34 at the default medium setting, against 17 for Haiku 4.5 with reasoning. The honest caveats are the price tier and the tokenizer: the headline rate only holds below 100K tokens, and the new tokenizer counts about 30% more tokens than Haiku 4.5, so Anthropic itself settles the real saving at roughly 75% rather than the advertised 90%.

Sources

Analysis generated: 2026-10-09