The full text is 1479 words in total, and the estimated reading is 6 minutes.
Today, Anthropic, the world's leading AI technology company, unveiled the next generation of Claude 3 models, marking a new era in AI cognitive capabilities.
Spanning the Claude 3 Haiku, Claude 3 Sonnet, and the flagship Claude 3 Opus, each model redefines the industry standard with its unique level of intelligence and performance.
A new benchmark in intelligence: the Claude 3 OPUS is at its peak
OPUS, the pinnacle of the Claude 3 series, has demonstrated amazing strength over its peers in a variety of AI system evaluation benchmarks, such as Undergraduate Expert Knowledge (MMLU), Graduate Level Expert Reasoning (GPQA), and Basic Mathematics (GSM8K).
In particular, OPUS has risen to the forefront of general intelligence by demonstrating near-human comprehension and fluency in complex tasks.
Claude 3 model series.
The Claude 3 model has significantly improved in terms of analysis, refined content creation, and generation, and has shown a higher level of communication in non-English languages such as Spanish, Japanese, and French. In a comparison with the top existing models, the Claude 3 has established a leading position in several dimensions.
Instantaneous response, endless possibilities
Each Claude 3 model is impressively responsive in real-time, enabling near-instant processing of customer service chats, autocomplete, or data extraction tasks.
Haiku, as the fastest and most cost-effective intelligent model on the market, can quickly digest complex studies including charts and graphs in as little as three seconds**;
Sonnet, on the other hand, maintains high speed while improving the level of intelligence, especially suitable for scenarios such as knowledge retrieval and sales automation.
Although the OPUS is similar in speed to the Claude 2 series, it is in a league of its own thanks to a higher level of intelligence**.
Visual power and accurate memory
Not only is the Claude 3 series a breakthrough in language processing, but it also demonstrates its extraordinary ability in visual comprehension, capable of handling a wide range of visual elements such as **, charts, graphs, and technical charts. This is significant for organizations with large knowledge bases stored in non-text form, unlocking new ways of processing information.
The Claude 3 series also successfully addresses the shortcomings of its predecessor in understanding and responding to context, significantly reducing unnecessary rejection. The data shows that Opus, Sonnet, and Haiku have significantly fewer refusal to answer when faced with system guardrail prompts, reflecting more refined comprehension and higher situational adaptability.
Accuracy upgrades, equal responsibility
Claude 3 opus is more accurate than Claude 2 in answering complex factual questions1 This is more than a twofold increase while reducing the percentage of incorrect answers. In the future, the Claude 3 model will also introduce a citation function, which enables the model to cite sentences in specific documents to verify answers, thereby improving the credibility of the output information.
In addition, the Claude 3 series excels at handling long form and contextual information, capable of handling contextual windows up to 200k, and is expected to scale up to 1 million token inputs for specific customer needs. Among them, OPUS achieved an accuracy rate of more than 99% in the "Needle in a Haystack" (NIAH) assessment, demonstrating excellent retrieval and recall capabilities.
Innovate responsibly and securely
Anthropic has always adhered to the principles of responsible design in the creation of the Claude 3 collection. The company has set up a dedicated team to reduce model false positives, resist the spread of bad information, prevent biological abuse, maintain election fairness, and limit the ability to replicate autonomously, and use constitutional AI methods to enhance the security and transparency of models.
Claude 3 has made significant progress in reducing bias, outperforming the legacy model on the Question Answer Bias Benchmark (BBQ) and remaining at AI Security Level 2 (ASL-2), with a very low potential catastrophic risk as assessed internally and by White House-mandated red teams.
The Claude 3 models have also been greatly improved in terms of ease of use, better able to follow complex instructions, respect brand guidelines, provide users with services that are closer to the human dialogue experience, and can more easily generate structured output, which can help with diverse application scenarios such as natural language processing and sentiment analysis.
Currently, Claude 3 Sonnet and OPUS are officially available in 159 countries and regions around the world through the Claude API, and Haiku will soon be added to the lineup.
claude 3 opus api**。
Sonnet is now ClaudeThe core support of the free experience on the AI platform, while OPUS provides services for Claude Pro subscribers.
claude 3 sonnet api**。
In addition, Sonnet is already available through Amazon Bedrock and is available in private preview on Google Cloud's Vertex AI Model Garden, followed by Claude 3 Haiku.
claude 3 haiku api**。
Anthropic will continue to iterate on the Claude 3 series of models in the future, and they also plan to introduce a number of enhancements, including tool usage and interactive coding, to meet the needs of enterprise and large-scale deployments. At the same time, they will stay on the line of security, ensuring that every leap in performance is accompanied by an upgrade in security measures.
Hotspot Engine Program