Summary
- Anthropic introduced Claude Haiku 5.5, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, significantly lower than Haiku 4.5's rates of $1 and $5.
- In benchmark tests, Haiku 5.5 achieved a score of 72.4% on the OSWorld 2.1 offline subset and 39.2% on Terminal-Bench 4.0, outperforming OpenAI's GPT-6 Luna but trailing behind Sonnet 5.5's 70.6% in coding assessments.
- Additionally, Anthropic reduced the cache-read cost for Sonnet 5.5 to $0.10 per million tokens and will be providing monthly API credits for Max and Team subscribers this week.
On Wednesday, Anthropic launched Claude Haiku 5.5, which it claims is its most economical and rapid small model to date. This model is designed for tasks that require high throughput, such as summarizing documents, querying databases, and providing live customer support.
For developers who integrate Claude into their applications, the cost is set at $0.10 per million input tokens and $0.50 per million output tokens for prompts that do not exceed 100,000 tokens. Tokens, which are small pieces of text roughly equivalent to three-quarters of a word, are the billing unit for AI services. This pricing aligns with that of OpenAI's GPT-6 Luna, which launched on September 22.
Haiku 4.5 had a higher rate of $1 for input tokens and $5 for output tokens, making the new pricing 90% lower. For prompts exceeding 100,000 tokens, users can expect a 50% discount. Given that around 90% of requests to the previous model were below this threshold, Anthropic estimates an average saving of approximately 75% with Haiku 5.5.
The model is particularly tailored for handling repetitive, straightforward tasks such as customer support interactions and summarizing lengthy emails.
Despite its focus on efficiency, benchmarks indicate that Haiku 5.5 is quite capable. In the OSWorld 2.1 test, which evaluates an AI's ability to perform complex, multi-step tasks on a real computer, it achieved a success rate of 72.4%, surpassing OpenAI’s GPT Luna, which scored 48.9%.
In Terminal-Bench 4.0, which assesses the performance of AI agents in executing professional tasks, Haiku 5.5 scored 39.2%, compared to 16.4% for OpenAI's Luna and 0% for Haiku 4.5.
For context, Anthropic’s Claude Sonnet 5.5 achieved a score of 70.6% in similar evaluations.
In a test with a simple logic question, Haiku 5.5 responded almost instantly, but it provided an incorrect answer, highlighting the need for caution in fully relying on its outputs.
On the GDPval-AA v2.1 scale, which evaluates models based on their performance across 44 professions using an Elo rating system, Haiku 5.5 received a score of 1620, while Luna scored 1437 and Haiku 4.5 received 735.
This is the first Haiku model to feature an adjustable effort setting, allowing users to balance cost against the quality of responses. Anthropic has also lowered Sonnet 5.5's cache-read price to $0.10 per million tokens, which applies to previously processed text.
Haiku 5.5 follows closely on the heels of the launch of Opus 5.5 on September 22 and Sonnet 5.5 on September 28, completing the suite of Claude 5.5 models that Anthropic had promised. The release of Opus 5.5 came shortly after CEO Dario Amodei's call for the industry to moderate advancements in AI capabilities.
Currently, Haiku 5.5 is accessible on the Claude website, as well as through Amazon Web Services, Google Cloud, and Microsoft Azure under the name claude-haiku-5-5. Anthropic is also introducing monthly API credits this week: $100 for Max 5x plan subscribers, $200 for Max 20x plan subscribers, and up to $500 for Team plans, shared among users.