Anthropic Drops Claude Haiku 5.5: 75 Percent Cheaper and Built for Subagents
Anthropic just released Haiku 5.5. It costs 75 percent less than 4.5, runs faster, and is built to act as a cheap subagent for your app.
Models · Source: Hacker News
What happened
Anthropic just launched Claude Haiku 5.5. It is their fastest and cheapest small model to date. They designed it specifically for high-volume, repetitive tasks like database queries, compactions, and classification requests. The model is available right now on all major platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can also access it directly through the Claude Platform.
The pricing structure is incredibly aggressive. Haiku 5.5 costs about 75 percent less to run than Haiku 4.5 on average. Input tokens are just ten cents per million for prompts under 100,000 tokens. Output tokens are fifty cents per million. Anthropic notes that 90 percent of requests fall under this token limit. They also introduced a brand new adjustable effort setting. This means developers can finally dial the model between optimizing for strict cost savings or higher intelligence based on the specific task at hand.
Anthropic did not stop at updating Haiku. They also cut the price of cache reads for Sonnet 5.5 by a massive 50 percent. Cache reads now cost just ten cents per million tokens. This makes Sonnet 5.5 about 20 percent cheaper for most agentic workflows. To push adoption further, they rolled out free monthly API credits. Max 5x users get 100 dollars, Max 20x users get 200 dollars, and Team subscribers get up to 500 dollars pooled across their accounts.
Key facts
- 75% — Average cost reduction compared to Haiku 4.5
- $0.10 — Cost per million input tokens for prompts under 100k tokens
- 90% — Percentage of previous Haiku requests under 100k tokens
- 72.4% — Haiku 5.5 score on OSWorld 2.1 offline subset
- $500 — Maximum monthly API credit pool for Team subscribers
Why it matters
This completely changes the math for multi-agent systems. You no longer need to burn expensive Sonnet or Opus tokens on basic data extraction or simple summaries. You can use Haiku 5.5 as a highly capable, dirt-cheap sidekick. It pulls the raw data, formats it perfectly, and hands it off to a larger model for the heavy lifting. Companies like Cognition and Rogo are already doing exactly this in production. Cognition uses Haiku 5.5 as a sidekick in Devin Fusion to cut costs while maintaining top-tier coding scores.
The second-order effect is a massive drop in latency for user-facing AI applications. Asana tested the model for their AI Teammates product and reported a 30 percent reduction in latency for task completions. They also saw up to 2.5 times faster inference per agent turn. If you build live customer support bots or browser automation tools, speed is your absolute main bottleneck. Haiku 5.5 removes that friction entirely while simultaneously slashing your monthly API bill.
For builders
Offload basic tasks to cheap subagents
Stop using Sonnet for simple classification or summaries. Route those high-volume tasks directly to Haiku 5.5 instead. You will save a massive amount of money and speed up your application for end users.
Leverage cheaper Sonnet cache reads
Anthropic halved the cost of Sonnet 5.5 cache reads to ten cents per million tokens. If you run complex agentic loops with heavy context, your profit margins just improved by 20 percent. You can either pass these savings to your users or keep the extra profit for your startup.
Build browser automation tools right now
Anthropic updated their Python and TypeScript SDKs to support computer and browser use in beta. Haiku 5.5 is fast enough to make these real-time interactions feel incredibly snappy. Builders should start creating autonomous web scrapers and QA bots today.
My take
Anthropic is commoditizing the bottom layer of intelligence. By making Haiku 5.5 dirt cheap and incredibly fast, they are forcing every builder to adopt multi-agent architectures. I love this aggressive move because it makes building complex AI products in public actually affordable for bootstrapped founders.
Original reporting: Hacker News. This is my rewrite and opinion.