The price of AI just fell off a cliff. Again. This time, though, the drop landed exactly where it helps the most: the small, fast models doing the invisible work inside your apps.
On October 7, Anthropic released Claude Haiku 5.5, and the headline number is hard to overstate. For prompts up to 100,000 tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens. The previous generation, Haiku 4.5, charged $1.00 and $5.00. That is a 90 percent cut, and it puts Haiku 5.5 at exactly the short-context price of OpenAI’s GPT-6 Luna.
What you actually get
Haiku has always been the workhorse of the Claude family. It skips the showpiece demos and does the summaries, the document Q&A, the classifications, the subagent tasks, the thousand little jobs that run every hour inside customer support systems and browser tools. Speed and cost are its whole job, and this release goes all in on both.
Three other changes are worth knowing. Haiku gets adjustable effort levels for the first time, from Low to Max, so developers can dial thinking depth up or down. The same announcement quietly halved Sonnet 5.5’s cache-read fees, a cost cut for the bigger sibling that compounds at scale. And Max and Team subscribers now receive $100 to $500 a month in API credits, a direct subsidy for the builders Anthropic wants to keep.
One quirk: the 90 percent discount applies below 100,000 prompt tokens. Above that line, rates rise to $0.50 input and $2.50 output, a 50 percent cut rather than 90. Anthropic says about 90 percent of Haiku 4.5 requests fell under the threshold anyway and estimates workloads will run roughly 75 percent cheaper on average. Worth keeping in mind if your documents are long.
The real battleground moved downstream
Two years ago, the race was about the smartest model. Now it is about the smartest cheap model, because the money in AI is increasingly made by agents, and agents burn tokens relentlessly. Every read, every summary, every plan, every check is a model call, and most of those calls are the boring kind: short prompts, simple jobs, repeated hundreds of times a day. A model at $1.00 per million tokens makes that affordable. A model at $0.10 makes it a rounding error. The gap between those two numbers decides how ambitious the agents get.
That is why matching GPT-6 Luna’s price matters. OpenAI and Anthropic have both decided the next wave of adoption will be fought at the subagent layer, the swarm of small, fast calls powering agentic workflows. Cisco is embedding Claude agents inside Webex meetings, and the economics of that only work if the small model is practically free. Now it is. The same cost pressure is rippling through physical AI too, from Singapore’s new school for training robots to the driverless rollout taking shape in London.
What builders should do
Re-price your workloads now, against the short-context tier, and watch for the 100K-token cliff if your prompts are long. Try the effort dial: adjustable thinking depth on a small model is new, and if Low effort handles your classification tasks at this price, that is a second discount hiding inside the first. And remember that cheaper tokens do not fix bad architecture. Still, a 90 percent cut on the workhorse tier is the kind of move that unlocks categories nobody has tried yet, because they could not afford to try.
Be First to Comment