On October 7, Anthropic made a move that should make every developer’s budget spreadsheet look friendlier: it released Claude Haiku 5.5 at roughly a tenth of its predecessor’s price. For prompts up to 100,000 tokens, the model costs $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 charged $1.00 and $5.00. That’s a 90 percent cut — and it lands Haiku 5.5 at exactly the same short-context price as OpenAI’s GPT-6 Luna.
Anthropic calls it the cheapest, fastest, and most capable small model it has shipped. The marketing is what it is. What’s actually interesting is what the move says about where the real fight in AI has moved.
The workhorse tier
Haiku has always been the workhorse of the Claude family. It doesn’t do the showpiece demos. It 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. For that work, speed and cost are the entire job, and this release goes all-in on both.
Beyond the headline price, three changes are worth knowing. First, Haiku gets adjustable effort levels, from Low to Max, so developers can dial thinking depth up or down depending on whether they need an instant answer or a careful one. Second, the same announcement quietly halved Sonnet 5.5’s cache-read fees — a cost cut for the bigger sibling that compounds at scale. Third, Max and Team subscribers now get $100 to $500 a month in API credits, a direct subsidy for the builders Anthropic wants to keep around.
One quirk in the pricing: 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.
Why cheap small models are the whole game now
Two years ago, the competition was about who could build the smartest model. Now it’s about who can make the smartest cheap model — because the money in AI is increasingly made by agents, and agents burn tokens relentlessly.
Think about what an AI agent actually does. It reads a page. Summarizes it. Plans a step. Reads another page. Checks its work. Every step is a model call, and most of them are the boring kind: short prompts, simple jobs, repeated hundreds of times a day. At $1.00 per million tokens, that’s affordable. At $0.10, it’s a rounding error. The difference between those two numbers decides how ambitious the agents get.
That’s why matching GPT-6 Luna’s price matters. OpenAI and Anthropic have both decided the battleground for the next wave of adoption is the subagent layer — the swarm of small, fast calls powering agentic workflows. Enterprises are embedding agents inside the meetings and inboxes where work already happens, and the economics of that only work if the small model is practically free. Now it is.
What builders should do this week
Re-price your workloads
Compare your current Haiku 4.5 or GPT-6 Luna spend against Haiku 5.5’s short-context tier. Watch for the 100K-token cliff if your prompts run long — the 5x markup above the threshold makes cost planning awkward for document-heavy pipelines.
Try the effort dial
Adjustable thinking depth on a small model is new, and it might matter more than the price. Many agent workloads over-think simple steps because the model has no way to be told “just answer.” If Low effort handles your classification tasks at this price, that’s a second discount hiding inside the first.
Don’t forget the architecture
Cheaper tokens don’t fix bad architecture. But zoom out: prices are falling at every layer of the AI stack at once, from the chips to the networking to the tokens. The trend is one-directional.
There’s a cynical read — the labs are racing each other to the bottom, and someone will eventually raise prices. Maybe. But computing history usually runs the other way: cheaper compute doesn’t destroy the business, it explodes the number of things people build with it. Agents that read every support ticket in real time. Assistants that summarize every meeting without being asked. Tools that classify, translate, and double-check as a matter of course instead of as a premium feature.
The optimistic version is simple: AI gets cheaper, the quiet background jobs get smarter, and the bill stays flat. For once, the price war is happening exactly where users benefit most.
Be First to Comment