What happened
Anthropic released Claude Haiku 5.5, the newest small model in the Claude family. The company calls it the cheapest, fastest and most capable small model it has shipped.
That 75% is Anthropic's estimate across real workloads, not a flat sticker cut.
VentureBeat reports a steeper per-token move: a 90% price cut for requests under 100,000 tokens. Input starts at $0.10 per million tokens, output at $0.50 per million.
VentureBeat says that matches OpenAI's GPT-6 Luna. The 75% and the 90% measure different things, so do not mash them together.
On speed, The Verge reports that only Anthropic's Opus models in Fast Mode outrun Haiku 5.5. It is also the first Haiku-class model with an adjustable effort level.
The target jobs are the repetitive ones: summaries, compactions, database queries and classification requests.
Anthropic also pitches it as a coding subagent working under Opus 5.5 or Sonnet 5.5, and for jobs where speed matters, like live customer support and browser use.
There is a side dish. Anthropic is halving Sonnet 5.5 cache reads, from $0.20 to $0.10 per million tokens. It says that makes Sonnet 5.5 around 20% cheaper on most agentic work.
Max and Team subscribers also get pocket money, per VentureBeat: a monthly API credit where Max 5x gets $100, Max 20x gets $200, and Team plans share up to $500.
Benchmark comparisons against Haiku 4.5, GPT-6 Luna and Sonnet 5.5 are vendor-reported. Treat them as a brochure until someone independent runs the same tests.
It is available now on all platforms.
Why it matters
Small models do the unglamorous bulk of the work. Nobody writes a victory post about classifying a million support tickets.
But somebody pays for it. That somebody just got a discount.
When the price of the dull stuff falls this far, the math on what is worth automating changes. Tasks that were too small to bother with start to look like free lunch.
The expensive brain plans. The cheap model does the legwork.
Anthropic is openly positioning the small model as the worker bee under Opus and Sonnet. A very corporate org chart, except nobody has a birthday.
And if VentureBeat's numbers hold, the price match with GPT-6 Luna turns the small-model aisle into a straight price war. Good for buyers. Less good for anyone selling a thin wrapper on a pricey model.
What to do with it
- Re-price your workloads.Test Haiku 5.5 on a sample of your summary or classification jobs; 75% is an average, not a promise.
- Mind the 100,000-token line.The reported 90% cut covers requests below it, so long-context jobs may save less.
- Play with the effort dial.Turn it down for easy calls and up for hard ones instead of overpaying for everything.
- Go tiered.Route routine steps to Haiku 5.5, keep Sonnet or Opus for planning, and on Max or Team, spend the new API credit testing it.
Ciao, sapiens. The small robots will take it from here.


