Anthropic launched, on this Wednesday (7), the Claude Haiku 5.5, a new artificial intelligence (AI) model from the company aimed at tasks that require speed, high volume processing, and low cost.
The launch updates the Haiku 4.5 version and completes the new Claude 5.5 family, which also includes the models Opus 5.5 and Sonnet 5.5. According to Anthropic, the Haiku 5.5 is their fastest, cheapest, and most capable small model so far.
The company states that the new model costs, on average, 75% less to operate than the Haiku 4.5. For requests up to 100,000 tokens, however, the price reduction per token reaches 90%.
Claude Haiku 5.5 targets repetitive tasks
- Anthropic developed the Haiku 5.5 for applications that need to process large volumes of requests without necessarily requiring the company's larger models;
- Among the cited uses are summarization, context compression, database queries, and classification. The model was also designed for tasks sensitive to speed, such as real-time customer service and web navigation;
- Another possibility is to use it as a subagent alongside larger models, such as Sonnet 5.5 and Opus 5.5, especially in programming tasks;
- The company claims that the Haiku 5.5 is its fastest model so far in standard speed.
Price drops to R$ 0.53 per million tokens
The cost is one of the main new features. For prompts up to 100,000 tokens, Haiku 5.5 costs US$ 0.10 per million input tokens (about R$ 0.53) and US$ 0.50 per million output tokens (about R$ 2.65).
Above 100,000 tokens, the prices rise to US$ 0.50 (about R$ 2.65) per million input tokens and US$ 2.50 (about R$ 13.25) per million output tokens.
Anthropic states that about 90% of requests made to Haiku 4.5 were within the 100,000 token limit. Therefore, the largest price reduction should affect most uses of the previous model.
The company also halved the cache reading price of Sonnet 5.5, from US$ 0.20 to US$ 0.10 per million tokens. According to Anthropic, this reduces the model's cost by about 20% in most agent tasks.
Claude Haiku 5.5 is also the first model in the Haiku category to offer a system of adjustable levels of effort. The configuration allows choosing how much reasoning the model should use before responding. Thus, developers can prioritize speed and cost in simple tasks or increase effort for work that requires more capacity.
Anthropic presents this possibility as a way to balance cost and intelligence according to each application.
The model also has a context window of one million tokens, according to the platform's documentation.
Performance increases in computer and programming tasks
Anthropic also presented its own test results to show the performance jump compared to Haiku 4.5.
In OSWorld 2.1, a benchmark that measures the ability of agents to operate a computer to perform complex tasks, Haiku 5.5 achieved 72.4%, compared to 15.7% of Haiku 4.5.
In Terminal-Bench 4.0, focused on terminal programming tasks, the new model reached 39.2%, while Haiku 4.5 had recorded 0%.
In the test Humanity’s Last Exam, Haiku 5.5 obtained 45.9% without tools and 57.4% with tools, compared to 10.2% and 18.7%, respectively, of the previous model.
The numbers are results disclosed by Anthropic itself and do not represent an independent evaluation of the model.
The Claude Haiku 5.5 is also the first model in the Haiku category to offer an adjustable system of effort levels – Image: Samuel Boivin/Shutterstock
Read more:
- US investigates OpenAI and Anthropic after incidents with AI agents How much does the CEO of one of the most important AI companies today earn
- Anthropic releases the most powerful AI for teams that hunt for vulnerabilities before hackers
Comparison with OpenAI models
Anthropic also placed Haiku 5.5 alongside OpenAI models in its performance charts.
For example, in the OSWorld 2.1, Haiku 5.5 achieved 72.4%, while GPT-6 Luna appears with 48.9%. The numbers are 39.2% and 16.4%, respectively, in the Terminal-Bench 4.0.
As the results were selected and published by Anthropic itself, they should be interpreted within the methodology used by the company, and not as an independent comparison between the models.
Anthropic emphasizes security
The company also claims to have improved the security behavior of Haiku 5.5 compared to Haiku 4.5.
According to Anthropic, the new model presented fewer episodes of misaligned behavior and less willingness to collaborate with misuse.
In the area of cybersecurity, the protections are more strict than in Haiku 4.5, but less restrictive than those applied to the company's more recent models. The system allows a broader range of defensive tasks, but blocks intrusion tests and other techniques considered more associated with attacks.
Biological safeguards follow the same pattern used in Sonnet 5, Sonnet 5.5, and Opus 5, according to the company.
API credits are available to subscribers
Along with the launch, Anthropic announced monthly API credits for subscribers of the Max and Team plans.
Max 5x subscribers will receive US$ 100 (approximately R$530) in credits per month, while Max 20x subscribers will have US$ 200 (approximately R$1,060). For Team customers, the benefit reaches up to US$ 500 (approximately R$2,650), shared among the organization's users.
Credits can be used with any Anthropic model on the platform and were created to encourage the development of applications and agents that use the API.
The company also updated its Python and TypeScript SDKs, which now offer beta support for computer and browser use.
Haiku 5.5 is already available
The Claude Haiku 5.5 is already available on Anthropic's platforms, including Amazon Web Services (AWS), Google Cloud, and Microsoft Azure. On the Claude Platform, developers can use the model identified as claude-haiku-5-5.
With the launch, Anthropic completes the update of its main line Claude 5.5 in just over two weeks. Haiku 5.5 arrives as the most economical option in the family, while Sonnet and Opus remain directed towards more complex tasks.
Rodrigo Mozelli is a journalist graduated from the University of São Paulo Methodist (UMESP) and, currently, is a writer for Olhar Digital.
