Anthropic Launches Claude Haiku 5.5 With Token Prices Cut by Up to 90%
Fundacion Rapala – Anthropic has introduced Claude Haiku 5.5, a new AI model designed to handle large volumes of simple and repetitive work at a much lower cost. Released on October 7, 2026, the model targets tasks such as document summarization, information classification, database queries, and smaller jobs performed by AI agents. However, the biggest attraction is its pricing. For prompts below 100,000 tokens, Anthropic has reduced token prices by up to 90 percent compared with Claude Haiku 4.5. This shift could matter for developers and businesses that process thousands or even millions of requests every day. Instead of using an advanced and expensive model for every task, companies can assign routine work to Haiku 5.5. In practice, that approach could lower operating costs while allowing larger Claude models to focus on jobs that demand deeper reasoning, stronger analysis, or more complex decision-making.
Token Prices Drop Sharply for Prompts Under 100,000 Tokens
The new pricing structure makes Claude Haiku 5.5 particularly attractive for shorter prompts. Anthropic prices the model at $0.10 per one million input tokens and $0.50 per one million output tokens when prompts stay below 100,000 tokens. By comparison, Claude Haiku 4.5 costs $1 per one million input tokens and $5 per one million output tokens. Therefore, the difference reaches 90 percent for eligible workloads. Longer prompts receive a different rate. For requests above the 100,000-token threshold, Haiku 5.5 costs $0.50 per one million input tokens and $2.50 per one million output tokens. That still represents a 50 percent reduction from the previous generation. Anthropic says roughly 90 percent of Haiku 4.5 requests fall below the 100,000-token limit. As a result, the lowest pricing tier could apply to a large share of everyday Haiku workloads.
Read More : What Is a Firewall? Meet the Digital Guard Protecting Your Devices
Anthropic Expects Average Workload Costs to Fall by 75 Percent
Lower token rates do not always translate directly into identical savings because newer models can consume tokens differently. Anthropic says Claude Haiku 5.5 uses a new tokenizer, which changes how text is divided and processed. After accounting for those differences, the company estimates that running workloads on the new model should cost about 75 percent less on average than using Haiku 4.5. That figure could become important for organizations where AI usage has moved beyond occasional experiments. Customer service systems, document processing tools, internal assistants, and automated workflows can generate huge numbers of requests. Even a small difference in the cost of each request can become significant at scale. Haiku 5.5 therefore appears aimed at a growing problem in enterprise AI: finding enough intelligence for routine work without paying premium-model prices. For developers, lower costs may also make it easier to test new AI features before deploying them more widely.
Haiku 5.5 Becomes a Supporting Player for Powerful AI Agents
Anthropic is positioning Claude Haiku 5.5 as a supporting model that can work alongside more capable models such as Claude Sonnet and Opus. The idea resembles a team in which complicated assignments go to senior specialists while smaller jobs are delegated to faster, cheaper workers. For example, a larger model could build the structure and analysis for a financial presentation. Haiku 5.5 could then retrieve specific revenue figures needed for individual slides. This division of labor can become especially valuable for AI agents, which may complete dozens of smaller actions before finishing one larger assignment. Using an expensive model for every step can quickly increase costs. Instead, developers can route straightforward tasks to Haiku while reserving stronger models for difficult reasoning. If implemented carefully, this approach could make sophisticated AI agents more affordable without requiring companies to use the least expensive model for every part of a workflow.
Read More : 140 Million Indonesians Watch YouTube Every Month as Viewing Habits Shift
Anthropic Reports Major Performance Gains Over Haiku 4.5
Lower pricing would matter less if performance declined sharply, so Anthropic is also emphasizing benchmark improvements. According to the company’s reported tests, Claude Haiku 5.5 scored 1,620 on GDPval-AA v2.1, compared with 735 for Haiku 4.5. It also reached 72.4 percent on OSWorld 2.1, while its predecessor recorded 15.7 percent. On Terminal-Bench 4.0, the new model reached 39.2 percent under the test configuration reported by Anthropic. FrontierCode 1.1 produced a score of 46.4 percent. These figures suggest a substantial generational improvement across several types of tasks. However, readers should interpret vendor benchmarks carefully. The results were reported by Anthropic rather than independently verified in the information provided. Moreover, Claude Sonnet 5.5 remained ahead of Haiku 5.5 across all listed benchmarks. Haiku therefore appears designed around efficiency and value rather than replacing Anthropic’s more capable models.
A New Effort Setting Lets Developers Balance Cost and Reasoning
Claude Haiku 5.5 also introduces an effort setting that controls how much reasoning the model performs before producing an answer. Anthropic uses medium effort as the default, creating a balance between performance and computational demand. Increasing the setting can improve results on difficult tasks, although stronger reasoning may require more resources. The difference becomes visible in Anthropic’s reported Terminal-Bench results. Haiku 5.5 achieved roughly 39 percent at maximum effort, while performance at medium effort was around 20 percent. That gap highlights an important point for developers: headline benchmark numbers may not always represent the default experience. Still, adjustable effort provides useful flexibility. A simple classification request may not require extensive reasoning, while a coding or multi-step agent task could benefit from a higher setting. Instead of treating every request equally, developers can choose how much computational effort makes sense for each workload and potentially manage costs more carefully.
Cheaper AI Models Could Change How Businesses Build Automation
Claude Haiku 5.5 reflects a broader shift in AI development from simply building more powerful models toward making intelligence affordable at scale. Many business workflows do not require the strongest model available. They require something fast, reliable, and inexpensive enough to run repeatedly. A 90 percent token-price reduction for shorter prompts could make that calculation very different for companies processing large volumes of routine tasks. However, price should not become the only consideration. Organizations still need to evaluate accuracy, latency, privacy requirements, reliability, and real-world performance before moving important workflows to a new model. Independent testing will also help clarify how closely Anthropic’s benchmark claims match everyday use. Even so, Haiku 5.5 presents an interesting proposition: instead of asking one expensive AI model to handle everything, developers can build systems where different models take different responsibilities, potentially making advanced automation more practical and financially sustainable.