On Thursday, June 20, OpenAI competitor Anthropic released its most powerful AI model to date, Claude 3.5 Sonnet.

In multiple performance tests covering reading, programming, math, and vision, Claude 3.5 Sonnet outperformed a host of other AI models, including GPT-4o, the company's flagship model Claude 3 Opus, while some tests may not fully reflect the use of AI in reality due to their niche nature and special cases.

Anthropic's product manager, Michael Gerstenhaber, confidently stated, "For businesses, what matters is whether AI can help them meet their business needs, not whether AI is competitive in performance testing. From this perspective, I believe Claude 3.5 Sonnet will surpass any other product we currently have, as well as any other product in the industry."

In terms of pricing, the new model is priced the same as the previous 3 Sonnet model, which is $3 per million input tokens and $15 per million generated tokens, with a context window of 200,000 tokens, equivalent to about 150,000 words. Token is the subdivision unit of data, such as "fan," "tas," and "tic" in the word "fantastic."

Users can now try the new model for free through Anthropic's web client and iOS app, while users who subscribe to Claude Pro and Claude Team will enjoy a five-fold rate limit. In addition, the new model is also available on Anthropic's API as well as hosted platforms such as Amazon Bedrock and Google Cloud's Vertex AI.

Later this year, the company will also release larger and better models, such as Claude 3.5 Haiku and Claude 3.5 Opus, the latter of which will feature functions such as network search and preference memory.

The strongest visual model, doubling the speed, adding humor, and content iteration functionality

Compared with the previous model, Claude 3 Opus, the new model has achieved improvements in many aspects of performance.

For example, compared with 3 Opus, Claude 3.5 Sonnet is better at understanding complex instructions and subtle differences. It can even better grasp humor concepts, despite AI's usually poor performance in humor.

For applications that require fast response, such as customer service chatbots, 3.5 Sonnet's processing speed is twice as fast as Claude 3 Opus, and its cost is only one-fifth of the latter.

In terms of visual analysis, 3.5 Sonnet can interpret charts and graphics more accurately and transcribe text from "imperfect" images with distortion and visual artifacts.

In addition to the new model, Anthropic has also launched the new feature "Artifacts Preview," which is a workspace that allows users to edit and iterate on the content generated by AI models.

Imagine you are using an AI assistant to help you write code. When you make a request to the AI, it generates a segment of code. In Artifacts, this segment of code not only appears for display but also in a form that can be manipulated and modified, like a "artifact" or a "draft."

Next, you can iterate on this code, meaning you can modify it, add new functionality, or communicate with AI assistant "Claude" to tell it your modified ideas or new requirements. Based on your feedback, AI will generate code again, and you can continue this process until the code meets your expectations and can be practically implemented.

This process is like you and AI working together to constantly refine and improve the final product. Artifacts provides a platform for you to interact with AI models more easily and effectively manage and optimize the generated content.

Currently, Artifacts is in preview, and Anthropic plans to add new features in the future, such as support for large-team collaboration and knowledge-base storage.

In addition, while Claude 3.5 Sonnet is an advanced AI model, it is not perfect and may still make mistakes, according to media reports. Nevertheless, its capabilities may be enough to attract developers and businesses to turn to Anthropic's platform. After all, this is the most important thing for Anthropic.

The improvement in model part is attributed to training data, but the source is unclear.

Anthropic's product manager, Michael Gerstenhaber, said the improvements were due to adjustments to the model architecture and new training data, including AI-generated data. Gerstenhaber did not reveal specific information about which data brought about these enhancements.

For the sake of protecting trade secrets and avoiding legal challenges, the specific details of the training data have not been disclosed. However, like the company's previous AI models, Claude 3.5 Sonnet was trained on large amounts of text and images and attempted to "align" the model with users' intentions based on feedback from human testers, in an attempt to prevent the model from generating harmful or problematic text.

At present, the court has not made a ruling on whether suppliers such as Anthropic and their competitors (including OpenAI, Google, Amazon, etc.) have the right to use public data (including copyrighted data) for training without providing compensation to the creators of these data.

What does the new model mean for Anthropic and the AI ecosystem?

The media pointed out that in the continuous evolution of AI technology, Anthropic's Claude 3.5 Sonnet model, while not bringing about revolutionary changes, represents the current reality of AI model development: while waiting for major scientific research breakthroughs, there is a continuous performance improvement, that is, a step-by-step approach.

In recent months, flagship products including Google's Gemini 1.5 Pro and OpenAI's GPT-4o have only achieved small improvements in benchmark testing and performance. Due to the limitations of existing model architectures and the enormous computing resources required for training, the industry has not witnessed another huge leap like the one from GPT-3 to GPT-4.

As generative AI suppliers shift their focus to data preparation and licensing, investors are becoming more cautious about the expected return on generative AI investments. Thanks to the favorable position of Amazon (and Google's minor support), Anthropic is relatively unaffected by this pressure. However, Anthropic's expected revenue is only slightly below $1 billion by the end of 2024, still lagging behind OpenAI.

Anthropic has a growing customer base including well-known brands such as Bridgewater, Brave, Slack, and DuckDuckGo, but there is still room for improvement in corporate reputation. It is worth noting that PwC recently chose to work with OpenAI rather than Anthropic to resell generative AI products to enterprises.

Anthropic realizes that building an ecosystem around models (rather than isolated models) is the key to retaining customers as the gap between model capabilities narrows.

On the one hand, in order to retain customers, Anthropic is strengthening tool development. These tools include Artifacts mentioned earlier, which allow developers to have deeper control over the internal functionality of AI models and allow AI models to perform specific operations in applications.

On the other hand, the company is expanding its team and market. Anthropic hired Instagram co-founder as product manager, indicating the company's emphasis on product development. In addition, the company has offices in London and Dublin, which helps to expand the market coverage of its products.

"When you are building an application, end users do not need to care about the model or optimization details used behind the scenes, and engineers can use tools to optimize the experience, of which cost is an important factor," said Michael Gerstenhaber, Anthropic's product manager.

Editor/Somer