What Haiku 5.5 is for
Haiku 5.5 is built for the work you do in bulk and don't want to overpay for. Anthropic is pointing it at jobs like classification, summarising, pulling data out of documents, live customer support, voice agents and in-app assistants, the kind of work where being quick and cheap beats being brilliant.
It's made to grind through repetitive tasks at scale, things like database queries and mass summaries. It can also sit under the bigger models and do sub-agent work for Opus 5.5 or Sonnet 5.5 on coding and other heavier jobs.
What it costs
The real story is the price. For prompts under 100,000 tokens, Haiku 5.5 starts at $0.10 per million input tokens and $0.50 per million output. Anthropic reckons that's about 75% cheaper than the last Haiku, version 4.5.
Run longer prompts and it climbs, to $0.50 per million input and $2.50 per million output. There's also an effort dial you can turn to trade speed and smarts against cost. The model takes up to a million tokens of context and can write outputs as long as 128,000 tokens.
A cybersecurity safeguard, new to Haiku
One thing is genuinely new here. Anthropic says this is the first Haiku-class model with built-in safeguards for a narrow set of high-risk cybersecurity requests. Those guardrails target the specific cases and, the company says, leave normal day-to-day use alone.
It's a small signal that even the cheap workhorse models are now getting the same safety treatment the bigger ones already have.
The IPO behind it
The IPO is the backdrop to all of it. Haiku 5.5 lands as Anthropic speeds up product work ahead of a planned listing, widening the lineup and chasing more enterprise customers as rivals pile in.
The launch came with a couple of sweeteners too. Anthropic is cutting cache-read prices on Sonnet 5.5, and handing monthly API credits to Claude Max and Team subscribers. Both make it cheaper and easier for developers to build on Claude.
So here's a company making its platform stickier right before it asks public investors to buy in, and it's doing that while plenty of people are loudly arguing the whole AI boom is overpriced.