Anthropic released Claude 3 Haiku on March 13, 2024, positioning it as a fast and cost-efficient option for enterprise use. The company says Haiku combines advanced vision capabilities with competitive performance on industry benchmarks.
Performance and speed
According to Anthropic, Claude 3 Haiku is three times faster than peer models for the vast majority of workloads, ingesting about 21,000 tokens per second for prompts under 32,000 tokens (roughly 30 pages). This speed supports batch processing, running many small tasks in parallel, and responsive chat experiences.
Anthropic notes that prompts exceeding 32,000 tokens may be ingested 30–60% more slowly; the company expects improvements in the coming weeks. Image processing can introduce additional latency.
Cost model and examples
Haiku uses a 1:5 input-to-output token pricing ratio, which Anthropic designed for enterprise workloads that often require longer prompts. The company gives concrete cost examples: for approximately one US dollar, Claude 3 Haiku can process and analyze about 400 U.S. Supreme Court cases (each case estimated at 10,000 tokens) or roughly 2,500 images (each image estimated at 1.6K tokens).
Security and robustness
Anthropic emphasizes enterprise-grade security and robustness for Haiku. The company performs rigorous testing to reduce harmful outputs and model jailbreaks. Additional safeguards include continuous system monitoring, endpoint hardening, secure coding practices, strong data encryption protocols, and strict access controls. Anthropic also conducts regular security audits and engages experienced penetration testers to proactively find and remediate vulnerabilities.
Further technical and safety details are available in the Claude 3 model card.
Availability
Starting March 13, 2024, customers can access Claude 3 Haiku via the Claude API and on claude.ai for Claude Pro subscribers. The model is already available on Amazon Bedrock and is expected to arrive soon on Google Cloud Vertex AI.
Footnotes
- Prompts containing over 32,000 tokens may experience 30–60% slower ingestion speeds.
- Anthropic estimates each U.S. Supreme Court case at about 10,000 tokens.
- Each image is estimated at roughly 1.6K tokens.
