Anthropic on Wednesday released Claude Haiku 5.5, a small artificial intelligence model the company says is its cheapest, fastest and most capable in that class to date.
The company said the new model costs about 75% less to run on average than its predecessor, Haiku 4.5. Anthropic is aiming it at high-volume, cost-sensitive work such as summaries, database queries, classification and live customer support.
Along with the launch, Anthropic said it is halving the price of cache reads on its midsize Sonnet 5.5 model. The company also said it will begin offering monthly API credits to some paid subscribers.
For prompts of up to 100,000 tokens, Haiku 5.5 costs 10 cents per million input tokens and 50 cents per million output tokens, according to the company’s pricing table. Haiku 4.5 costs $1 and $5 for the same amounts.
Prices rise for longer prompts, to 50 cents for input and $2.50 for output per million tokens. Anthropic said about 90% of requests to the previous Haiku model fell under the 100,000-token threshold.
The company said its 75% savings estimate accounts for a new tokenizer, which means Haiku 5.5 uses slightly more tokens to finish a given task.
Anthropic reported large gains over Haiku 4.5 across its own benchmark results. On OSWorld 2.1, a test of how well AI agents operate a computer to complete multistep tasks, Haiku 5.5 scored 72.4% on an offline subset, compared with 15.7% for Haiku 4.5 and 48.9% for OpenAI’s GPT-6 Luna.
On the Terminal-Bench 4.0 coding test, Haiku 5.5 scored 39.2%. Haiku 4.5 scored zero and Sonnet 5.5 scored 70.6%.
Anthropic said its larger Sonnet 5.5 and Opus 5.5 models remain better choices for complex coding work. It said Haiku 5.5 is best suited to narrower jobs that were previously too expensive to run, such as summarization or working as a subagent under a larger model.
Haiku 5.5 is also the first Haiku-class model with an adjustable effort setting, the company said. That lets users choose between lower cost and stronger performance.
The company described the model as its fastest yet at standard speed. It noted Haiku 5.5 still runs slower than its Opus models when those use Fast Mode.
Anthropic said early customers reported results in line with its testing. Aaron Vinh, a staff software engineer at Asana, said in a statement released by Anthropic that his team saw latency on task completions fall by more than 30% compared with the model the company uses now.
“It’s a noticeably snappier experience.”
Anthropic said Haiku 5.5 showed major improvements on almost all of its alignment evaluations compared with Haiku 4.5, including fewer cases of misaligned behavior and less willingness to help with misuse.
The company said the new model’s cybersecurity safeguards are stricter than those for Haiku 4.5 but somewhat looser than those on its other recent models. They allow a wider range of defensive security tasks than Sonnet 5.5’s safeguards, but still block penetration testing and other techniques more likely to be used by attackers, Anthropic said.
Starting Wednesday, Sonnet 5.5 cache reads cost 10 cents per million tokens, down from 20 cents. Anthropic said the change lowers the cost of Sonnet 5.5 on most agentic tasks by about 20%.
This week, the company said, it will roll out monthly API credits to Max and Team subscribers. Max 5x users will receive $100 a month and Max 20x users $200, while Team subscribers will get up to $500 pooled across their users. The credits can be used on any of the company’s models.
Anthropic also said it is adding beta support for computer use and browser use to its Python and TypeScript software development kits.
Haiku 5.5 is available now on all of Anthropic’s platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, the company said.
