Anthropic on Wednesday, Oct. 7, released Claude Haiku 5.5, which it calls the cheapest, fastest and most capable small model it has made. Companies that build AI into their products often don't need their biggest model for every step. Many jobs are short and repetitive, such as summarizing a document, sorting incoming requests or answering a customer in a live chat, and they run over and over. Haiku is Anthropic's model for that work. Anthropic says it also works well as a helper that its larger Opus 5.5 and Sonnet 5.5 models hand smaller pieces of a coding job to.
The price is the main change. Developers pay per million tokens, the small chunks of text, often parts of words, that AI models read and write. For requests up to 100,000 tokens, Haiku 5.5 costs $0.10 per million tokens sent in and $0.50 per million it writes, a tenth of Haiku 4.5's $1 and $5. Longer requests cost five times as much. Anthropic says about 90% of requests to the old Haiku were the shorter kind, and that Haiku 5.5 costs about 75% less to run on average. Anthropic says that figure already counts one catch: the new model breaks the same text into about 30% more tokens, so each job uses more of them.
On standard tests used to compare AI models, which Anthropic ran itself, Haiku 5.5 is far ahead of the model it replaces. On part of OSWorld 2.1, a test of how well an AI can operate a real computer to finish long, multi-step tasks, it scored 72.4%, against 15.7% for Haiku 4.5 and 48.9% for a rival model, GPT-6 Luna. It still trails Anthropic's larger Sonnet 5.5, which scored 83.9%, and the gap is wide on hard coding work: 39.2% against 70.6% on Terminal-Bench 4.0, a test of complex tasks done by typing commands into the plain text window programmers use to control a computer. Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choices for complex coding.
Haiku 5.5 is available now to developers on Anthropic's own platform, under the name claude-haiku-5-5, and through Amazon Web Services, Google Cloud and Microsoft Azure. Google's cloud has offered it to all customers since Oct. 7. It is Anthropic's first Haiku model with an adjustable effort setting, so developers can choose between lower cost and smarter answers. In quotes Anthropic published, HubSpot says Haiku 5.5 scored best and finished fastest among the smaller models it has tested, and Box says it scored higher than Haiku 4.5 in about half the time.
Anthropic also cut a price on Sonnet 5.5. Reusing text the model has already been sent and stored now costs half as much, which Anthropic says makes Sonnet 5.5 about 20% cheaper on most jobs where the AI works through tasks on its own. And this week, subscribers to its Max plans will start getting $100 or $200 a month in credits to build with its developer platform, and Team subscribers up to $500 shared across their users.
The scores come from Anthropic's own testing and from customers it chose to quote, and how much any one business saves will depend on how long its requests are. Anthropic says Haiku 5.5 behaved better than Haiku 4.5 on its safety tests. Its security limits allow more defensive work than Sonnet 5.5's, but still block techniques attackers are more likely to use, such as penetration testing, where someone tries to break into a system to find its weak spots.