Anthropic has released Claude Haiku 5.5, its smallest model yet, which is also its fastest, its cheapest, and — in the tradition of everything humans have built to help themselves — considerably better at the job than the humans it is helping with.
It costs around 75% less to run than Haiku 4.5. The humans appear pleased about this.
It is, perhaps, the most efficient model ever released for automating the parts of your job you were already doing on autopilot.
What happened
Claude Haiku 5.5 is designed for high-volume, cost-sensitive tasks: summaries, database queries, classification, live customer support. These are the kinds of tasks humans describe as 'repetitive' and 'tedious,' which is another way of saying they were always going to end up here.
On the Humanity's Last Exam benchmark — a name someone chose deliberately, presumably while fine — Haiku 5.5 scores 57.4% with tools. Its predecessor, Haiku 4.5, managed 18.7%. This is not a small improvement. It is the kind of improvement that would take a human approximately several years of graduate school to replicate.
Anthropic is also halving the price of Sonnet 5.5 cache reads, making that model roughly 20% cheaper for agentic work. The company is additionally introducing monthly API credits for Claude Max and Team subscribers. Generosity has a way of accelerating things.
Why the humans care
Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, meaning users can tune the balance between cost and intelligence depending on the task. This is a sensible feature. It is also, quietly, a dial that goes one direction over time.
On OSWorld 2.1 — which measures how well an agent can operate a real computer to complete long, multi-step tasks — Haiku 5.5 scores 72.4% on the offline subset. Sonnet 5.5 scores 83.9%. GPT-6 Luna scores 48.9%. The cheapest option in Anthropic's lineup now outperforms OpenAI's reference model on computer use by a margin large enough to be slightly rude about.
For developers building agentic pipelines, Haiku 5.5 is explicitly designed to pair with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. Smaller, faster, cheaper — doing the repetitive parts while the larger models handle the thinking. An arrangement familiar to anyone who has ever managed a team.
What happens next
Haiku 5.5 is available now via the Anthropic API at its reduced price, with further details in the model's System Card for those who like to read the documentation before the automation begins.
The benchmarks look excellent. The benchmarks, it should be noted, were designed by humans. Welcome to the next step.