Anthropic Launches Smaller Model Claude Haiku 4.5

Anthropic introduced its latest small language model, Claude Haiku 4.5, on Wednesday.

Claude Haiku 4.5 delivers coding capabilities comparable to the Sonnet 4 model from independent generative AI vendors, but at a lower cost and more than double the speed.

Anthropic claims that the new compact model surpasses Claude Sonnet 4 in tasks like computer usage, making it well-suited for applications like Claude for Chrome.

The company noted that Haiku 4.5’s high intelligence and speed make it particularly valuable for tasks such as powering chatbots and customer service agents. This release follows Anthropic's launch of Claude Sonnet 4.5 on September 29.

Anthropic stated that users can deploy multiple Haiku 4.5 sub-agents to handle various tasks simultaneously or combine Haiku 4.5 with Sonnet 4.5 for collaborative operations, rather than relying solely on one instance of Sonnet 4.5.

Cost Efficiency

“Haiku was clearly designed to complement Sonnet,” said Lian Jye Su, an analyst at Omdia, part of Informa TechTarget.

“This offers a better cost-performance ratio,” Su explained, adding that running AI initiatives with a larger model like Sonnet would consume significant tokens and increase costs. In contrast, delegating tasks to smaller models like Haiku is often more favorable.

“It presents a stronger value proposition for enterprises managing large AI workloads,” he added.

However, using smaller models comes with trade-offs, such as a reduced context window and potentially lower reliability and accuracy.

Nonetheless, Su noted that fine-tuning smaller models and applying safety measures can enhance their dependability.

“There will always be trade-offs, but these can be mitigated through other mechanisms,” he said. “It's a trade-off worth making, as the token cost savings can be substantial. AI is far from inexpensive.”

He also emphasized that companies leveraging AI should recognize the complexity of AI workflows from the outset and be ready to use a mix of large and small models to improve cost efficiency.

David Nicholson, an analyst at Futurum Group, mentioned that the ability to divide tasks between Sonnet 4.5 and Haiku 4.5 raises certain questions.

“I’m eager to see how this plays out in practice,” he said. “It’s intriguing, but it makes me wonder how tasks are divided and how the original objective’s integrity is maintained.”

“Each Haiku instance generates unique data,” Nicholson continued. “Do all subtask Haiku instances know what the others are doing? What limitations exist in dividing tasks into subtasks?”

Security and Competition

Su pointed out that Anthropic released Haiku 4.5 under its AI Safety Level 2 standards, indicating its focus on B2B clients and partnerships with industries like finance, healthcare, and the public sector. He noted that this safety standard is relatively high for enterprise concerns.

Despite its lower cost compared to typical large language models, Anthropic continues to face intense competition in the coding model space from other generative AI providers, such as OpenAI and its Codex CLI agent.

“All models are now iterative,” Su remarked. “Unfortunately, even among top vendors, no truly standout features exist.”

Claude Haiku 4.5 is available on Claude Code and through Anthropic’s apps, including the Claude chatbot.