Claude Haiku 4.5 – a new lightweight AI model launched by Anthropic

What is Claude Haiku 4.5？

Claude Haiku 4.5 is Anthropic’s latest high-performance, cost-efficient small AI model. It delivers coding performance comparable to the previous flagship model Claude Sonnet 4, and even surpasses it in some tasks—while costing only one-third as much and running over twice as fast. The model offers excellent alignment and safety, rated at AI Safety Level 2 (ASL-2). Claude Haiku 4.5 can be deployed via the Claude API, Amazon Bedrock, and Google Cloud Vertex AI, making it ideal for real-time, low-latency use cases such as chat assistants, customer service agents, and programming companions.

Main Features of Claude Haiku 4.5

• Coding Capabilities: Performs exceptionally well in programming tasks, supporting multiple languages and generating high-quality code. Ideal for rapid prototyping and multi-agent collaboration projects.

• Real-Time Interaction: Optimized for low-latency tasks such as chatbots, customer support agents, and pair programming, ensuring fast and smooth user experiences.

• Multi-Task Processing: Works alongside the flagship Claude Sonnet 4.5 to decompose complex problems into smaller sub-tasks for parallel execution, greatly improving efficiency.

• High Security: Extensively tested for safety and alignment, with a very low incidence of harmful behavior—one of Anthropic’s safest models to date.

• Cost Efficiency: Priced at only one-third of Claude Sonnet 4—$1 per million input tokens and $5 per million output tokens, offering exceptional value.

Performance Overview

Coding Ability:

• SWE-bench Verified: Achieved 73.3% accuracy, comparable to Claude Sonnet 4 (72.7%), approaching top industry levels.

• Terminal-Bench: Scored 41.0%, outperforming Claude Sonnet 4 (36.4%).

• OSWorld: Scored 50.7%, significantly higher than Claude Sonnet 4 (42.2%), showing strong performance in computer-use tasks.

Mathematical Ability:

• With Python Tool Support: Achieved 96.3% accuracy, demonstrating excellent performance.

• Without Tools: Scored 80.7% accuracy, slightly lower but still exceeding many large-scale models.

Multilingual Ability:

• MMMLU Benchmark: Achieved an average accuracy of 73.3% across 14 non-English languages, indicating strong multilingual understanding.

Official Project Page

• Website: https://www.anthropic.com/news/claude-haiku-4-5

Application Scenarios

• Programming Assistant: Provides developers with code generation, debugging suggestions, and rapid prototyping support—ideal for collaborative, multi-agent projects.

• Chat Assistant: Builds real-time, responsive chatbots that handle user queries smoothly and efficiently.

• Customer Support Agent: Assists customer service teams in quickly answering inquiries and delivering accurate, consistent information.

• Pair Programming: Supports programmers with real-time feedback, logic optimization, and coding suggestions to improve code quality.

• Educational Tutoring: Helps students learn programming by explaining complex concepts, generating exercises, and providing detailed answers.