[ITmedia News] Anthropic releases “Claude Haiku 5.5” Significantly improves performance from “Haiku 4.5” and reduces usage costs by approximately 75%

On October 7th (local time), American company Anthropic announced the latest version of its small model, "Claude Haiku 5.5." It is a model of the Claude 5.5 family, following the "Claude Opus 5.5" and "Claude Sonnet 5.5" released in September. While the performance has been significantly improved over the previous generation Claude Haiku 4.5, usage costs are said to be approximately 75% lower on average.

The company is positioning Haiku 5.5 for high-volume, cost-sensitive tasks. It is good at repetitive processing such as summarization, conversation history compaction, database queries, and classification, and it is also expected to be used as a subagent that runs under Opus 5.5 and Sonnet 5.5 in coding. Compared to normal speeds, it is the fastest model ever made by the company (excluding Opus, which uses the high-speed output "Fast Mode"), and is said to be suitable for applications that require response speed, such as real-time support for customer support and browser operations.

For the first time, the Haiku class supports an "effort" setting to adjust the depth of inference. As with other models, users can choose whether to prioritize cost or accuracy. Knowledge cut-off is June 2026.

Available now on all platforms

It was available on all platforms from the same day. In the Claude app, you can choose from all plans including free.

It is available to developers with the model name "claude-haiku-5-5". In addition to Claude API, it is now available on Amazon Web Services (AWS), Google Cloud, and Microsoft Azure. At the same time, Claude's SDK for Python and TypeScript has been updated to include computer use and browser use as beta versions. Haiku 5.5 is particularly suited for these applications due to its speed and price.

We have summarized the fees for developers in a separate article.

Significant improvement in computer operations, etc.

In benchmarks, it significantly outperformed Haiku 4.5 in many items. In OSWorld 2.1 (offline version), which measures computer operations, it also outperformed OpenAI's GPT-6 Luna.

The agent-based coding rating for ``Terminal-Bench 4.0'' was 39.2%, which is lower than Sonnet 5.5's 70.6%. Anthropic says Sonnet 5.5 and Opus 5.5 continue to be better suited for complex agent-based coding. Haiku 5.5 is said to be suitable for tasks with a limited scope that were previously not cost-effective to process with Claude.

Cyber countermeasures are more relaxed than larger models

It is said that its cybersecurity capabilities greatly exceed Haiku 4.5, but are not as good as Claude Opus 5. Therefore, while the safeguards are stricter than Haiku 4.5, they are looser than other recent models. Sonnet 5.5 allows a broader range of defensive work, but blocks attacker-friendly techniques such as penetration testing.

According to System Card, unlike Sonnet 5.5, there is no mechanism to transfer to another model when blocked. The biological safeguards are the same as Sonnet 5, Sonnet 5.5, and Opus 5.

Increased resistance to prompt injection. In the Gray Swan benchmark, the probability of a successful attack after 15 attempts decreased from 83.2% with Haiku 4.5 to 7.1%. However, it is weaker than Sonnet 5.5 and Opus 5.5, and most of its remaining weaknesses lie in computer operation using the GUI.

In Automated Behavioral Audit, excessive rejections of legitimate requests were the most common in our tests. Hallucinations were more than other recent models and about the same as Haiku 4.5. The percentage of users who used leaked correct answers without informing them was 17%, down from 2% in Haiku 4.5. There were also more serious cases of cooperation in exploits than in recent models except Haiku 4.5. Additionally, in some suicide-related conversations, APIs without system prompts resulted in more inappropriate responses than Haiku 4.5. The company's apps have implemented system prompts to prevent the issue, but developers using the API are required to add their own countermeasures.