By Roman Aglaganov
Most AI headlines go to the biggest models. Most AI work, though, is done by small ones: the models that summarise, classify, look things up and answer support chats millions of times a day, where every fraction of a cent counts. On 7 October, Anthropic released the new top of that tier, Claude Haiku 5.5.
Cheaper, by a lot
For prompts up to 100,000 tokens, which Anthropic says make up about 90% of requests to its previous Haiku, the price is $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 cost $1 and $5. Longer prompts cost $0.50 and $2.50. Anthropic’s average estimate is that it costs about 75% less to run. Alongside it, Anthropic halved the price of cache reads for Sonnet 5.5 and added a monthly API credit for Max and Team subscribers.
And better
Haiku 5.5 is the first Haiku with an adjustable effort setting, so developers can trade cost for more careful answers. On Anthropic’s published benchmarks, the jumps over Haiku 4.5 are large: 72.4% versus 15.7% on an offline subset of OSWorld 2.1, a test of operating a real computer through multi-step tasks, and 39.2% versus 0% on Terminal-Bench 4.0, an agentic coding test. Anthropic also reports it ahead of OpenAI’s GPT-6 Luna on the tests where both were scored.
The usual caveat: these are the company’s own numbers on its chosen benchmarks. Independent results will follow, and customer workloads matter more than leaderboards.
Where it fits
Anthropic pitches it as a subagent for its bigger Opus 5.5 and Sonnet 5.5 models, the fast assistant that fetches a figure from a long report while the big model does the reasoning, and for speed-sensitive jobs like live customer support. Its cybersecurity safeguards are tighter than Haiku 4.5’s and still block penetration testing.
The bigger trend is that yesterday’s frontier keeps becoming today’s budget option. A small model that can operate a computer for a fraction of a cent changes which tasks are worth automating.
