The model arrived on October 7 with a price tag attached to the announcement, which is not how Anthropic usually introduces its products. Claude Haiku 5.5, the third model in the company’s 5.5 series, is aimed squarely at the cost-sensitive end of the market, and its token prices came in as much as 90% below the previous generation.
The timing is deliberate. Anthropic is preparing to go public, and Reuters framed the release inside that process, a company filling out its product line before it has to answer to public shareholders about growth. The pricing also sits almost on top of OpenAI’s cheapest tier, a direct challenge in the part of the market where most of the volume lives.
Haiku 5.5 keeps the million-token context window that has become the family’s signature, a capacity far larger than what most rivals offer. The change is in what it costs to use. Anthropic set input and output prices at $0.10 and $0.50 per million tokens for prompts up to 100,000 tokens, an average reduction of about 75% from Haiku 4.5. The company also cut the cached-read price for its Sonnet 5.5 model to $0.10.
The price cut is paired with a claim of better performance, and the numbers are striking. On OSWorld, a benchmark that measures how well a model operates a computer, Haiku 5.5 scored 72.4%, up from 15.7% for the previous version. On Terminal-Bench 4.0, a test of terminal programming, it rose from zero to 39.2%. On GDPval-AA, a knowledge-work benchmark, it reached 1,620, more than double the prior generation.
Those jumps are the kind of thing a company advertises when it wants investors to believe the product is getting better while it gets cheaper. The story Anthropic is telling is that capability and cost are not in tension, that each new model does more for less, and that the company can keep cutting prices without giving up its technological lead.
The context is a pricing war that has been running through the AI industry for two years. OpenAI cut prices repeatedly through 2024 and 2025, and the low end of the market has become a place where margins are thin and volume is everything. Anthropic, which has positioned itself at the premium end, is now reaching down into that market with Haiku.
The strategy has a clear commercial logic. Most of the queries that flow through AI systems are simple and repetitive, and they are won on price rather than on the quality of the reasoning. A model that can handle those tasks for a tenth of the old cost frees a company to win the volume business while reserving its more expensive models for harder work.
It also matters for the IPO. Anthropic’s prospectus has drawn scrutiny, and some analysts have questioned whether the company can turn its research lead into durable revenue. A cheaper model that expands the addressable market is one answer to that question, and the timing, days before a public listing, is unlikely to be a coincidence.
The competition is not standing still. OpenAI’s own low-cost tier keeps improving, and the two companies now track each other closely on price, a dynamic that favors customers but squeezes everyone who is not a clear leader. Anthropic is betting that its engineering can keep the gap in its favor.
For developers, the release is a straightforward win. The same class of work that used to cost a certain amount now costs less, and the model that does it is measurably better at the tasks developers actually assign it. That is the strongest case Anthropic can make for the release, and it does not need an IPO narrative to land.
The company has not said when it will list, and it has not confirmed the terms. But the sequence, a full product line, a price cut aimed at OpenAI, and a run of benchmark results, reads like a company putting its best face forward before the public gets its first real look at the financials.
Anthropic has built its lineup around a three-tier structure: Opus for the most demanding work, Sonnet for the mainstream, and Haiku for speed and price. The 5.5 series has rolled out across those tiers over recent months, and Haiku 5.5 completes the set. The company’s pitch to developers is that they can now move between tiers without changing their code, choosing the model that fits each task’s cost.
The benchmark results carry a subtext. The jump on OSWorld, from roughly 16% to 72%, suggests the cheap model can now do the agentic work of clicking through a screen that was once reserved for more expensive systems. That is the direction the whole industry is heading, toward models that act on computers rather than just answer questions, and it is a race in which price matters as much as skill.
The question that will follow Anthropic into its listing is whether a price war and a full product line translate into profit. Haiku 5.5 answers the product side of that question. The financial side will have to be answered in the prospectus.
