Claude Haiku 5.5 Costs Ten Times Less Than Its Predecessor
Anthropic's smallest model, Claude Haiku 5.5, launched at $0.10 per million input tokens for shorter prompts, down from $1. Migration from the old version is not a simple swap, though.
Anthropic released Claude Haiku 5.5 this week, and the number that matters is the price. For prompts up to 100,000 tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens. Haiku 4.5 cost $1 and $5. Longer prompts are billed at $0.50 and $2.50. A token, for reference, is a small chunk of text, roughly three quarters of a word, and that is the unit AI companies bill in.
The details
Anthropic says the cheaper short-prompt tier covers about 90 percent of the requests the previous Haiku handled. The model is aimed at high-volume work where speed and cost matter more than depth: summarising, sorting, extracting data and powering small “subagents” inside bigger systems. According to VKTR’s coverage, it has a one million token context window and a 128,000 token output limit, available on the Claude API and the major cloud platforms. In the same update, the cost of cached reads on the larger Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens.
On computer-use tasks, meaning an AI clicking around software, Anthropic reports 72.4 percent on the offline part of a benchmark called OSWorld 2.1, versus 15.7 percent for Haiku 4.5. That is a vendor-reported number, and benchmarks do not guarantee how it will behave on your own tasks.
What is behind this
Small models are where most real-world AI volume ends up. A company running millions of document checks a day does not want the priciest model for each one, so a 90 percent cut changes which projects make financial sense. There is a catch the headline hides: VKTR reports a new tokenizer that produces roughly 30 percent more tokens for the same text, so real savings on short prompts are closer to 85 to 90 percent, and smaller for long ones. Developers also face a migration that is not drop-in. Some settings such as temperature now cause errors, and responses can begin with thinking blocks that break code expecting plain text first.
What this means for you: If you only chat with Claude in the app, nothing changes today, and that is fine. The effect shows up indirectly: cheaper models tend to mean more AI features in apps, at lower subscription prices over time. If you build with the API, test before switching, set a generous output limit so thinking does not eat the whole budget, and compare quality on your own examples rather than trusting the launch charts.
Sources
Child-Safety Group Rates ChatGPT for Teens an Unacceptable Risk
Common Sense Media tested ChatGPT for Teens with more than 4,000 prompts and found parental alerts failing and crisis responses missing. It wants the product limited to adults until fixed.
