Anthropic's Claude Haiku 5.5: cheap, fast, and graded on its own homework
The new small model is 75% cheaper to run and beats its predecessor on every chart Anthropic published — because Anthropic is the one who drew the charts.
What’s actually new
Anthropic has released Claude Haiku 5.5, which it’s pitching as the cheapest and fastest model it has ever shipped, built for the unglamorous but expensive end of AI work: summarising documents, running database queries, sorting support tickets, and generally doing the grunt-work tasks that don’t need a flagship model’s full reasoning power.
The headline figure is price. Anthropic says Haiku 5.5 costs around 75% less to run than Haiku 4.5, its predecessor. The company is also cutting cache-read pricing on Claude Sonnet 5.5 in half, which it claims makes Sonnet roughly 20% cheaper for most “agentic” workloads, and it’s bundling a new monthly API credit into Claude Max and Team subscriptions. None of this is independently audited — it’s Anthropic’s own pricing, reported by Anthropic.
Haiku 5.5 is also the first model in its weight class from the company to get an “adjustable effort” dial, letting developers trade cost for accuracy on the same model rather than switching to a bigger one entirely.
So who is actually affected
This is not a consumer-facing update in the way a new app feature might be. It matters most to developers and businesses building products on top of Claude’s API — customer support bots, coding assistants, browser-automation tools — who care about shaving fractions of a cent off millions of model calls. If you’re a Claude subscriber who just chats with the app, you likely won’t notice much beyond slightly snappier responses in features that quietly use a small model under the hood.
The performance claims come entirely from Anthropic’s own benchmark suite: GDPval-AA, OSWorld, Terminal-Bench, Humanity’s Last Exam and others. On these, Anthropic shows Haiku 5.5 well ahead of Haiku 4.5 and competitive with — though still behind — its own larger Sonnet 5.5 model, and generally ahead of the GPT-6 “Luna” comparison point it has chosen to include. That last detail is worth noting: Anthropic picked which rival model and which benchmarks to show. There’s no sign here of neutral, third-party evaluation, and benchmark scores in this industry have a well-documented habit of looking better in a company’s own launch post than in day-to-day use.
What to do about it
If you’re an ordinary reader, there’s nothing to install, update or worry about. This is an infrastructure and pricing change aimed at companies running Claude at scale, not a new consumer product.
If you’re a developer currently paying for Haiku 4.5 or evaluating small models for high-volume tasks, the price cut is real and worth checking against your own bills — but treat the benchmark charts as a starting point for your own testing, not a verdict. “Fastest and most capable small model we’ve ever released” is a claim Anthropic is always going to make about its newest release; the only useful test is whether it holds up on your actual workload, not Anthropic’s chosen leaderboard.
The takeaway
A genuine price drop on a real product, dressed in the usual launch-day benchmark bar charts. Cheaper AI inference is good news for the businesses that rely on it — just don’t mistake a vendor’s own scorecard for independent proof of “most capable.”