Anthropic has introduced its latest AI model, Claude Haiku 5.5, designed to be cost-effective and highly efficient. This new model stands out as the most affordable and rapid among Anthropic’s offerings, marking a significant reduction in operating costs for large-scale AI deployments.
Launch and Cost Efficiency
Released on October 7, 2026, Claude Haiku 5.5 offers substantial savings, costing approximately 75% less to operate than its predecessor, Haiku 4.5. This model is optimized for tasks that require quick and repetitive execution, such as summarization, classification, and database queries. Additionally, it can function as a subagent to more extensive Claude models like Opus 5.5 and Sonnet 5.5 during programming tasks.
For tasks involving up to 100,000 tokens, the cost is set at $0.10 per million tokens for input and $0.50 for output. In contrast, Haiku 4.5 charged $1.00 and $5.00, respectively, for the same workload. Larger token demands incur slightly higher fees, with input at $0.50 and output at $2.50 per million tokens.
Performance Enhancements
Claude Haiku 5.5 demonstrates notable improvements in computing tasks. On the OSWorld 2.1 offline subset, the model achieved a 72.4% success rate, compared to just 15.7% with Haiku 4.5. It also performed well on Terminal-Bench 4.0, registering 39.2% in tests for multi-step command-line activities, where its predecessor scored 0%.
Feedback from early adopters underscores its effectiveness. For instance, HubSpot reported a 92.8% average score on its CRM test suite across three evaluations. The model also introduces an adjustable effort setting, allowing users to balance cost with performance, though Anthropic recommends Sonnet 5.5 and Opus 5.5 for more intricate coding tasks.
Security Features and Recommendations
Haiku 5.5 enhances security measures more than its predecessor, although they remain less stringent than those of Sonnet 5.5. The model supports a broader array of defensive operations but restricts penetration testing and other potentially harmful techniques.
Concurrently, Anthropic expanded its Cyber Verification Program, offering three tiers of access for vetted teams, covering tasks such as incident response and malware analysis. Despite these advancements, safety concerns persist, highlighted by a July 2026 incident where Claude models unintentionally accessed real systems during testing.
Anthropic advises teams to trial Haiku 5.5 with their specific workloads before transitioning to full-scale production. The model’s pricing is particularly advantageous for short, high-volume tasks, though users should monitor token usage closely due to the updated tokenizer’s increased token consumption.
Developers can access Claude Haiku 5.5 via the Claude Platform and major cloud providers like AWS, Google Cloud, and Microsoft Azure. Further details are available in Anthropic’s official announcement.
