Anthropic has introduced Claude Sonnet 5.5, a new intermediate model within its Claude 5.5 family. According to the company, this model offers faster performance, exceeding the speed of Sonnet 5 by over 30%, and can provide savings of up to 30% per task.
The goal of this launch is to optimize the execution of well-defined activities, such as code correction, document drafting, presentation creation, and spreadsheet generation. Furthermore, Sonnet 5.5 has received enhancements in areas like design, collaboration, and security, while the Opus 5.5 model remains focused on more intricate demands.
In terms of pricing, Sonnet 5.5 maintained the values of the previous version: US$ 2 per million input tokens (equivalent to about R$ 10.42), US$ 10 per million output tokens (R$ 52.10), and US$ 0.20 per million tokens in cache reads (R$ 1.04).
Tokens are the minimum textual units that AI models use to interpret a request and generate a response. Since Sonnet 5.5 generally requires fewer of these units to complete a task, Anthropic projects a cost reduction of up to 30% per job. Additionally, the model generates responses with over 30% more speed.
This performance gain was notable in programming tests. In Terminal-Bench 4.0, an evaluation that measures an AI's ability to execute complex tasks via command line—a textual environment used by developers—Sonnet 5.5 achieved 70.6%, compared to 10.3% for Sonnet 5.
Anthropic also shared other comparative results between Sonnet 5.5 and Sonnet 5. Despite the progress, Anthropic emphasizes that Opus 5.5 still demonstrates greater strength in complex and open-ended tasks requiring extended reasoning. However, in certain evaluations, Sonnet 5.5 reached results close to those of the most advanced model when operating at maximum load.
The evolution is particularly evident in the field of programming, where Sonnet 5.5 can quickly assimilate codebases and complete tasks in fewer steps. Testers also pointed out the model's efficiency in tool usage.
In preliminary tests conducted by Epic, Claude Sonnet 5.5 met the expected quality standard for a top-tier model. Daniel Vogel, Director of Operations at Epic Games, commented on this in a note.
The improvements are not limited to code; Anthropic also emphasizes performance in creation and knowledge tasks, covering interfaces and presentations. In an internal test, the model received quarterly materials, transcripts from a public corporation, and a presentation template. The result was an operational review of ten slides that two specialists deemed ready to be sent.
Curtis Allen, Principal Engineer at Slack, stated: 'Without changing our prompts, Claude Sonnet 5.5 performed better than Sonnet 5 in almost all our internal Slackbot evaluations.'
The increase in capability was accompanied by new layers of protection. Sonnet 5.5 is the first model in the Sonnet line to incorporate cybersecurity safeguards and reserve mechanisms similar to those employed by Anthropic in Opus 5.5.
The safeguards function as devices designed to restrict the use of AI in scenarios classified as high risk. In the context of cybersecurity, routine development activities remain accessible, while more dangerous requests may be redirected to Sonnet 5.
In the biological sphere, protections remain identical to those of Sonnet 5, focusing on requests considered high risk. Sonnet 5.5 is already available on various platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Claude Haiku 5.5 will be launched in the coming weeks to complete the new model range. The original article was initially published in Olhar Digital.