Anthropic cuts Haiku's price by 75% to match OpenAI's smallest model

Anthropic described the reduction as an average of 75% ^[3]. That means that on a blended basis Haiku 4.5 cost about four times as much as its replacement ^[14], and individual rates did not all fall by the same amount. Above the 100,000-token line both input and output cost five times the short-prompt rate ^[15]. In either tier an output token costs five times an input token ^[16].

The cut can be read three ways. If Haiku 5.5 is cheaper to serve than Haiku 4.5, Anthropic can charge a quarter of the old price ^[14] and keep its margin per token. If it is not, the cut buys volume ahead of an IPO expected before Thanksgiving ^[8]. And if most of the company's revenue comes from long prompts or from Sonnet and Opus, the models it sells for higher-level coding and enterprise work ^[12], the cut touches a small share of sales.

The second reading is the one the price-war argument needs. The S-1 Reuters obtained puts 2025 sales at $4.6 billion, up 12 times on the year ^[9]. Against that, the $42 billion loss ^[10] works out to about $9.13 for every dollar of revenue ^[17]. Sales the year before were roughly $383 million ^[18].

In my view the evidence supports the buyer side of that argument and not yet the vendor side. Enterprise customers are looking for ways to cut costs after what Yahoo Finance called the brief "tokenmaxxing" era of unfettered spending ^[11]. Some are turning to open-weight models from US and Chinese developers ^[7]. Neither report includes Anthropic's gross margin or its cost to serve a token, so nothing published links the 2025 loss to Haiku pricing.

What Anthropic is giving up is list price on its cheapest tier in the months before it asks public investors for money ^[8]. The margin-squeeze thesis would be confirmed if the prospectus shows gross margin falling in the quarters after cuts like this one. It would be wrong if cost per token fell as fast as price.