In the ever-evolving realm of artificial intelligence, Anthropic is making waves with the introduction of two game-changing innovations: the enhanced Claude 3.5 Sonnet and the all-new Claude 3.5 Haiku. These models promise to elevate the AI experience with groundbreaking advancements in coding, speed, and versatility.
The spotlight on the Claude 3.5 Sonnet shines brightly as it emerges with significant upgrades, pushing the boundaries of agentic coding and tool use tasks. This AI model sets new standards, particularly in coding, by skyrocketing its performance on industry benchmarks—making it a standout leader over the previous models. It demonstrates improved capabilities on SWE-bench Verified and TAU-bench, setting a high bar for coding and tool use tasks in sectors like retail and airlines. Developers can now tap into its prowess while enjoying the same affordability and speed as its predecessor.
The feedback from pioneers using Claude 3.5 Sonnet has been overwhelmingly positive. Companies in various industries, including GitLab, Cognition, and The Browser Company, have noted the remarkable improvements in reasoning, planning, and software automation. This makes the Sonnet model an ideal choice for driving complex, multi-step development processes and workflows.
Introducing a new chapter in AI interaction, Claude 3.5 Sonnet comes with an experimental but intriguing feature—computer use. This public beta offers developers a sneak peek into the future of AI capabilities where Claude can mimic human interaction with computers. From moving cursors to typing and clicking, this feature aims to make AI as adept at using computers as we are. While still in its developmental phase, major players like Asana, Canva, Replit, and others are already exploring its potential to automate intricate tasks.
Anthropic is also unveiling Claude 3.5 Haiku, which is set to become a new darling in AI with its commendable speed and state-of-the-art performance. This model promises enhancements across all skill sets, particularly excelling in coding tasks and offering unmatched efficiency in processing huge volumes of data swiftly.
These strides represent a significant leap toward AI’s future, and Anthropic is keen on ensuring responsible use. Rigorous pre-deployment testing has been undertaken to evaluate potential risks, maintaining safety as a top priority.
As these powerful AI models hit the market, developers are invited to explore and contribute feedback, further enriching the development of AI. The arrival of Claude 3.5 Sonnet and Claude 3.5 Haiku opens up exciting new avenues in AI technology, promising to redefine how we interact with machines and set new standards for innovation. Whether you’re a developer delving into AI capabilities or a business looking to enhance operational efficiency, these models offer a world of potential waiting to be unlocked.Imagine a world where artificial intelligence can seamlessly operate computers just like humans do. This new skill, currently in public beta, marks a monumental leap in AI advancement. But why is this capability so important? In today’s world, much of our work is conducted through computers. Empowering AIs to interact with software just like a person could open up an exciting array of applications that were previously impossible with existing AI assistants.
In recent years, AI has achieved remarkable feats, such as performing complex logical reasoning and understanding images. Now, the focus shifts to the next frontier: computer use. Instead of relying on specially designed tools, AI models are being developed to use everyday software by simply following instructions.
Our journey to this breakthrough involved building on our previous work in tool use and multimodal AI. Computers require the ability to interpret screens—essentially, images—and reason through operations based on what’s displayed. By combining these skills, we trained an AI named Claude to not only understand what’s happening on a computer screen but also how to use the software tools available to complete tasks.
Notably, Claude’s ability to generalize its training from simple software like calculators and text editors to more complex applications was astonishing. With safety in mind, the training avoided internet access. Even more impressive was Claude’s ability to translate written prompts into logical actions and self-correct when encountering obstacles.
However, this breakthrough didn’t come easily. It was a process filled with trial and error, resembling the “idealized” version of AI research—constant iteration and testing until progress was made. The hard work paid off, and now Claude has set a new benchmark for AI that operates computers as humans do.
With new capabilities come fresh safety challenges. While this advancement mainly enhances AI’s ability to apply its current cognitive skills, concerns about potential threats are still a priority. Currently, Claude remains at AI Safety Level 2, meaning it doesn’t necessitate higher safety measures than what’s in place today. This provides an opportunity to address any safety issues early, rather than introducing these capabilities into future models with increased risks.
In anticipation of possible risks, our Trust & Safety teams have extensively analyzed the new models. One major concern is “prompt injection,” a form of cyberattack where malicious instructions can cause AIs to misbehave. Developers using Claude’s computer-use function must take precautions against such attacks, and we offer guidance on safeguarding against these vulnerabilities.
There’s also the risk of intentional misuse of Claude’s skills. Our teams have developed tools to monitor and mitigate such activities, especially with the forthcoming U.S. elections and the potential for misuse that could undermine public trust. We’ve implemented systems to track and moderate any election-related activities, ensuring Claude’s safe and responsible operation.
Regarding privacy, we maintain our standard policy: our models are not trained on user-submitted data, which includes screenshots provided to Claude.
This new skill represents a shift in AI development. Historically, AI was adapted to fit specific tools. Now, the aim is for the AI to adapt to the tools—using existing software environments just like anyone else would. While Claude’s current capabilities are still developing—often slow and prone to errors—it’s a promising start. For example, while creating demonstration videos, we observed glitches, such as accidentally stopping recordings or taking unscheduled detours to view scenic photos.
Despite these quirks, we’re confident that computer use will quickly become faster, more reliable, and more practical for everyday tasks. It will be more accessible, even for those lacking in-depth software development skills. As we advance, our researchers will collaborate with safety teams to ensure these new capabilities are both fruitful and secure.
We encourage developers participating in our public beta to share their feedback, so we can continue enhancing the utility and safety of this cutting-edge functionality. As we navigate this exciting development, we look forward to collaboratively shaping a future where AI and humans can work together more seamlessly than ever before.
