Anthropic Releases Claude Haiku 5.5, a New Model Up to 10x Cheaper Than Haiku 4.5

Anthropic has unveiled Claude Haiku 5.5, a new compact model in the Claude family built for high-volume tasks where speed and cost matter most. It is already available to developers and users, and for requests of up to 100,000 tokens it costs 10 times less than Haiku 4.5.

On the model’s official page, Anthropic calls Haiku 5.5 its fastest model in standard mode. The exception is the larger Opus models when using the special Fast Mode.

Its main use cases are summarization, classification, database queries, customer support, browser automation, and serving as a supporting agent for larger models.

Price drops to $0.10 per million input tokens

For prompts of up to 100,000 tokens, pricing is $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 cost $1 and $5, respectively.

For longer requests, prices rise to $0.50 and $2.50 — here the advantage over its predecessor is 50%.

Anthropic estimates an average cost reduction of about 75%, since roughly 90% of requests to the previous Haiku fell within the 100,000-token threshold. The company also warns that Haiku 5.5 uses an updated tokenizer, so comparing the price of a single token alone is not enough.

At the same time, the cost of cache reads for Claude Sonnet 5.5 was cut in half — from $0.20 to $0.10 per million tokens.

In Anthropic’s tests, Haiku 5.5 far outperformed its predecessor

According to Anthropic’s own benchmarks, the gains were especially noticeable in computer-use and agentic tasks.

On OSWorld 2.1, the score rose from 15.7% to 72.4%; on Terminal-Bench 4.0, from 0% to 39.2%; and on the Chartography visual reasoning test, from 6.4% to 46.4%.

In the table published by Anthropic, Haiku 5.5 also scored higher than GPT-6 Luna on all five tests where the company provided values for both models. For example, on OSWorld it scored 72.4% versus 48.9%, on Terminal-Bench 39.2% versus 16.4%, and on GDPval-AA 1620 versus 1437.

However, these are results from testing published by Anthropic, not an independent comparative review. The larger Claude Sonnet 5.5 remains noticeably stronger than Haiku on complex agentic and professional tasks.

Haiku 5.5 is Anthropic’s first model in its class with an adjustable effort level: developers can choose between minimum cost and deeper reasoning.

In preliminary tests, Asana reported reducing task latency by more than 30% and speeding up individual agent turns by up to 2.5 times. HubSpot, for its part, said Haiku 5.5 delivered the best result among the compact models the company tested in its CRM benchmark.

Claude Haiku 5.5 is already available under the identifier claude-haiku-5-5, as well as through Amazon Web Services, Google Cloud, and Microsoft Azure.