Anthropic has introduced Claude 3, a new set of AI models that includes Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus.

Claude 3 Opus, the most powerful version, is reported to outperform GPT-4 and other models on benchmarks like MMLU and HumanEval.

Key Features and Capabilities
Claude 3 offers a wide range of skills including advanced reasoning, basic math, data analysis, information extraction, forecasting, content creation, and code generation. It also supports multiple languages like Spanish, Japanese, and French. When compared to other models, Claude 3 Opus consistently ranks higher on various performance tests.

- Claude 3 Haiku is the quickest and most affordable version.
- Claude 3 Sonnet is twice as fast as earlier Claude models.
- Claude 3 Opus is as fast as Claude 2.1, but with significantly improved abilities.

These models support up to 200K tokens in context windows, and can be extended to handle up to 1 million tokens for selected customers. Claude 3 Opus performed extremely well in the Needle In A Haystack (NIAH) test, which measures a model's ability to recall information from large datasets and handle lengthy prompts effectively.

Additionally, these models have strong visual processing abilities, enabling them to understand and analyze photos, charts, and graphs.

Improvements in Understanding and Accuracy
Claude 3 models show better understanding of complex requests and are less likely to reject tasks. Claude 3 Opus is particularly good at answering factual questions accurately and reducing mistakes or "hallucinations." These models also produce more organized outputs, such as JSON objects, compared to the Claude 2 versions.