Haiku 5.5 Arrives | API Pricing Drops to One-Tenth
機械翻訳 / Machine-translated
A
AI Friends
4日前・18分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever struggled with AI API costs being too high to integrate into your service? On October 7, 2026, Anthropic released a new small model at one-tenth the price of its predecessor. This article explains how the pricing works and how to think about making the switch.
Claude Haiku 5.5 is a small model announced by Anthropic (the AI company behind Claude) on October 7, 2026. It is considered the highest-performing model in the Haiku series.
It is available via API (the mechanism that lets other software call AI) under the model name claude-haiku-5-5, and can also be accessed through AWS, Google Cloud, and Microsoft Foundry.
On Claude.ai, Haiku 5.5 is now available across plans from free to paid tiers. It also supports Claude Code.
The Haiku series is smaller, faster, and cheaper than the higher-tier Sonnet and Opus models. It is well suited for processing large volumes of text and generating quick chat responses.
Pricing is determined per one million tokens (the unit AI uses when reading and writing text).
According to the official Anthropic pricing page, when the prompt (the instructions and materials passed to the AI) is 100,000 tokens or fewer, the input cost is $0.10 and the output cost is $0.50 per million tokens.
Haiku 4.5 was priced at $1.00 for input and $5.00 for output — meaning both input and output are now one-tenth the price.
Assuming an exchange rate of 150 yen to the dollar, that works out to roughly ¥15 for input and ¥75 for output.
However, the lower price only applies to short prompts. Once you exceed 100,000 tokens, the price rises to $0.50 for input and $2.50 for output — half the price of Haiku 4.5.
Unlike other models from Claude 4.6 onward, Haiku 5.5 has a unit price that varies based on prompt length. Use cases where you pass in entire long documents will not see the "one-tenth" pricing.
That said, according to ITmedia AI+, approximately 90% of requests to Haiku 4.5 were under 100,000 tokens. So for most users, this represents a significant price cut.
For prompt cache reads (the mechanism that reuses the same instruction text), the cost is $0.01 per million tokens for short prompts. Using batch processing (a method of bundling non-urgent tasks together) cuts that by an additional 50%.
Anthropic estimates the average cost reduction from Haiku 4.5 to be approximately 75%. This accounts for the fact that the new tokenizer (the mechanism that converts text into tokens) slightly increases token counts for the same text.
On the same day, cache read pricing for Claude Sonnet 5.5 was also reduced — from $0.20 to $0.10 per million tokens, cutting it in half.
The base pricing of $2.00 for input and $10.00 for output remains unchanged. Only the cache read price has changed.
In agentic workflows, the same materials are often read repeatedly. As a result, Anthropic states that costs will decrease by approximately 20% in many cases. The actual savings depend on how much cache is being reused.
For example, consider a bot that reads a company internal manual and answers hundreds of questions per day. Since it re-reads the same manual every time, the cache pricing has a significant impact.
According to vendor-reported figures shared by VentureBeat (not independently verified), the gains over Haiku 4.5 are notable:
Meanwhile, Sonnet 5.5 scores higher than Haiku 5.5 on all of these benchmarks, recording 70.6% on Terminal-Bench. Complex coding remains the domain of Sonnet and Opus.
Another new feature is effort (adjustable reasoning depth) — a first for the Haiku series. You can choose between a quick and cheap response or a more deliberate one. The default is medium.
According to VentureBeat, Terminal-Bench scores are approximately 39% at maximum effort and approximately 20% at medium. Difficult tasks require a higher effort setting.
The following pricing per one million tokens was compiled by VentureBeat as of October 7, 2026:
At the lowest price tier, Haiku 5.5 is priced identically to GPT-6 Luna. In terms of performance, Haiku 5.5 outperforms GPT-6 Luna on the benchmarks Anthropic chose — though identical pricing does not necessarily mean identical performance.
Total real-world cost depends on more than just the unit price. The number of retries and accuracy of answers also matter. If a cheaper model requires three attempts, it may end up costing more than a single successful attempt with a more expensive model.
Haiku 5.5 is better suited as a subagent handling delegated tasks rather than as an orchestrator managing complex work. Sonnet and Opus are better for heavy planning, while Haiku excels at handling large volumes of simple tasks.
The barrier to integrating AI into business operations is lowering even for small and medium-sized businesses in Japan. Here are three scenarios to consider.
The first is handling inquiries for an online shop. If the job is simply sorting thousands of daily emails into categories like "shipping," "returns," and "other," a small model is more than sufficient. A tenfold reduction in unit price will directly affect your monthly bill.
The second is summarizing meeting minutes. If a transcript fits within 100,000 tokens, the lower short-prompt pricing applies directly.
The third is serving as a backend for AI agents like Claude Code. The official Anthropic page lists acting as a subagent — one that carries out tasks planned by a larger model — as a primary use case for Haiku.
There are also some caveats. Haiku 5.5 has stricter cybersecurity restrictions than Haiku 4.5. Penetration testing (the practice of probing systems for vulnerabilities) is blocked under the standard safety measures. If you need it for security work, you will need to apply for Anthropic's approval program.
Additionally, according to Forkast, Haiku 4.5 will be discontinued on October 15. If you are currently using Haiku 4.5, it is worth verifying your implementation sooner rather than later.
If you switch based solely on the pricing table, your actual bill may end up higher than expected. Check the following three points:
As a first step, it is recommended to switch just one of your simplest, highest-volume tasks to Haiku 5.5. A week of comparison should give you a clear picture of the cost and quality differences.
Q. Is Haiku 5.5 really one-tenth the price?
A. For prompts of 100,000 tokens or fewer, both input and output are one-tenth the price. Beyond 100,000 tokens, it drops to half price.
Q. How much cheaper is it on average?
A. Anthropic estimates an average cost reduction of approximately 75% compared to Haiku 4.5. This figure accounts for the slight increase in token counts from the new tokenizer.
Q. Did Sonnet 5.5 base pricing also drop?
A. No. Only the cache read pricing was reduced, from $0.20 to $0.10 per million tokens. The base pricing of $2.00 input and $10.00 output remains unchanged.
Q. Can Haiku 5.5 handle complex programming tasks?
A. It is not well suited for that. Terminal-Bench 4.0 scores are 39.2% for Haiku 5.5 and 70.6% for Sonnet 5.5. For complex coding, choose Sonnet or Opus.
Q. Can it be used for free?
A. On Claude.ai, Haiku 5.5 is available across Free, Pro, Max, Team, and Enterprise plans. The API is pay-as-you-go.
As a next action, choose one of the simplest, highest-volume AI tasks you currently run and try it with Haiku 5.5.
This article is a cross-post from AI Friends.
