Anthropic shipped Claude Haiku 5.5 on October 7, and at first glance it looks like the usual small-model refresh. Look closer and the interesting move is in the pricing, not the parameter count. This is the first Haiku that can dial its own reasoning effort up and down per task, yet it sits at roughly the floor of what Anthropic charged for the previous generation.
Start with the money, because that is the part that actually matters. For prompts up to one hundred thousand tokens, input costs 0.10 dollars per million tokens and output costs 0.50. Cross that line and the price jumps fivefold to 0.50 in and 2.50 out. Anthropic points out that about 90 percent of Haiku 4.5 requests fell under the threshold, so for most workloads the model is around 90 percent cheaper; blended across tiers, average running cost is down about 75 percent. On the same day the company also cut Sonnet 5.5 cache-read pricing from 0.20 to 0.10 per million tokens, shaving another 20 percent off most agent workloads.
But the 90 percent figure needs a caveat, and the caveat is the kind that shows up on your real bill. Haiku 5.5 ships with a new tokenizer that splits the same text into roughly 30 percent more tokens than the old one. So the unit price drops while the unit count quietly rises. On top of that, thinking output is billed as output tokens, and at max effort the model is genuinely verbose. Independent measurements put the real per-task saving closer to a quarter off Haiku 4.5 than nine tenths. The headline number is a best-case scenario, not your invoice. Before you budget, run your actual prompt lengths through the formula instead of trusting the marketing page.
Capability is genuinely up a tier. On one knowledge benchmark it scores 1620, on a business benchmark 1578, on the offline GUI test 72.4 percent versus 15.7 for Haiku 4.5, and on the coding benchmark at max effort it lands 39.2. It reads images and documents, carries a one-million-token context, and can emit up to 128,000 tokens. Worth noting: its list price is identical to the rival Luna model at 0.10 and 0.50, so the two labs are now elbow to elbow at the cheap entrance.
Put Haiku 5.5 next to Anthropic's own pricier tiers and the strategy is clear. The mid tier charges 2 dollars in and 10 out, so Haiku is about 95 percent cheaper; against the flagship the gap is about 98 percent. In other words, Anthropic has dragged its cheapest tier down to match the rival while still earning margin on the expensive tiers. That pattern is plainly aimed at call volume, not just flagship sales.
The practical change is that Haiku 5.5 finally works as a real sub-agent. You can let the flagship or mid tier do the heavy lift and offload dozens of cheap summarization, classification, and routing calls to Haiku. One expensive model orchestrating a swarm of cheap ones is now a buildable cost pattern. Take a concrete setup: a support system uses Haiku 5.5 for intent detection and ticket triage, and only genuinely hard cases escalate to a pricier model, which can pull total invocation cost down by an order of magnitude. This is also a clear signal from Anthropic as it heads into a likely public listing: they are competing hard for enterprise call volume.
Do not mistake it for a general replacement. On complex reasoning it still trails the flagship, and the moment a prompt crosses the threshold the price snaps back to frontier-model territory. It fits high-volume, latency-sensitive, tolerantly-imperfect work: support triage, content tagging, bulk extraction. It also exposes an effort parameter from lowest to highest, so the same model can do a cheap lookup or chew a harder step, but max-effort thinking is billed too, and the burn rate climbs visibly with the setting.
There is another angle worth stating. This price cut is not isolated. It is part of a broad downward move in small models, and when the cheapest tier can already handle summarization, classification, and triage, the old habit of throwing a flagship at every task stops making sense. The real saving is not picking the cheapest model. It is splitting work into tiers so that whatever a small model can do never reaches a flagship. Put differently, the value of Haiku 5.5 is not how strong it is. It is that tiering finally makes financial sense on the bill. For teams just starting out, that matters more than chasing benchmark scores.
Here is a worked example that makes the math concrete. Suppose you run ten thousand support triages a day, each with a two-thousand-token prompt and a three-hundred-token answer. On the old tier that is roughly thirty-five dollars a month. On the new tier the same volume drops to four or five dollars. But the moment your prompt crosses one hundred thousand tokens because context grew, the unit price quintuples and the monthly bill can jump past sixty dollars. So what actually decides cost is not the model name but how many of your requests sit comfortably under the threshold. Keeping long context short and using cache well matters far more than agonizing over which tier to pick.
What this means for you: if you run a lot of repetitive, cheap, fast AI calls, Haiku 5.5 is about the most economical execution layer available right now. Before you ship it, price it against your actual prompt lengths, and watch two things specifically: the roughly 30 percent extra tokens from the new tokenizer, and the fivefold jump once you cross 100,000 tokens. The discount is real. The average is not your bill.