10 Claude

Anthropic

Anthropic’s Claude represents a family of large language models (LLMs) specifically designed with safety, helpfulness, and honesty as core principles.^[1] Developed by Anthropic, a public benefit corporation founded by former OpenAI researchers, Claude distinguishes itself through its emphasis on constitutional AI training methods and robust safety measures.^[2] For users, Claude offers sophisticated reasoning capabilities, extensive context windows, and a commitment to accurate and well-sourced responses.

Claude’s development philosophy prioritises harmlessness and truthfulness alongside capability, resulting in an AI assistant that tends to be more cautious and transparent about limitations.^[3] The platform offers multiple model variants optimised for different use cases, from quick daily tasks to complex analysis requiring deep reasoning. Claude’s training emphasises a nuanced understanding of context, ethical reasoning, and the ability to engage with complex and multi-faceted problems. The models have gained recognition for their thoughtful responses and strong performance across analytical and creative tasks.

 Watch

Have a look at the newly redesigned user interface for Claude below.

Interface

Claude is accessible through Anthropic’s direct chat interface. Interact with this guide below to understand the various functionalities within the Claude interface.

Anthropic offers both free and paid versions of Claude. While using the free plan, Claude’s context window and daily message limit (i.e., the outputs) may vary based on demand. Users are notified if a prompt exceeds the available context window or if they reach the daily message limit. However, users on the ‘pro’ plan benefit from a context window of over 200,000 tokens (approximately 500 pages of text), with message limits five times greater than those of the free service. For relatively short conversations with Claude, users can send at least 45 messages every five hours. Once users reach the limit, it is reset every five hours.

Claude 4 Family of Models

Claude 4 family of models represents the company’s most powerful models.^[4] Claude Sonnet 4 is the company’s current standard model, which provides a balance of capability, speed, and cost-effectiveness. Whereas Claude Opus 4 is the company’s most capable model in the Claude family, designed for the most complex and challenging tasks. Claude Opus 4 and Sonnet 4 are hybrid models offering two modes: near-instant responses and extended thinking for deeper reasoning. After its introduction with Claude 3.7, each comes with ‘extended thinking’. The mode allows for step-by-step processing, which is visible to the end user. As part of this release, extending thinking comes with the use of tools. That is, both models can utilise tools like internet searching while they are ‘thinking’. The combination of these features means that the model can alternate between ‘reasoning’ and searching the internet to improve its responses.

Claude Sonnet 4 stats include:

• Context widow of 200,000 tokens
• Maximum output of 64,000 tokens
• The knowledge cutoff is March 2025

Claude Opus 4 stats include:

• Context widow of 200,000 tokens
• Maximum output of 32,000 tokens
• The knowledge cutoff is March 2025

Claude 3.7 Sonnet

Anthropic’s Claude 3.7 Sonnet model is a hybrid reasoning model. When it was announced, Anthropic described it as ‘an ordinary LLM and a reasoning model in one’.^[5] The model’s performance was described as being able to ‘understand nuanced instructions and context, recognise and correct its own mistakes, and create sophisticated analysis and insights from complex data.’^[6] The model with the ‘extended’ mode turned off is similar to previous Claude models, providing direct responses without showing internal reasoning. Stats include:

• Context widow of 200,000 tokens
• Maximum output of 8,192 tokens
• The knowledge cutoff is November 2024
• Available on all Claude plans—including Free, Pro, Team, and Enterprise
• Achieved 78.1% accuracy on LegalBench Benchmark testing 

Extended Thinking

The introduction of Claude 3.7 Sonnet came with the feature ‘extended thinking’. A Claude user can choose when the model responds normally or takes longer to think before replying. This option was added as a user interface option with a button to enable the ‘extended thinking’ mode (see the interface walkthrough above). Enabling ‘extended thinking’ is similar to the ‘reasoning’ models (discussed in more detail in Prompting a Reasoning Model). The mode allows for step-by-step processing, which is visible to the end user. Users can choose when they want the model to respond and when they prefer it to take a bit longer before answering.

Users can view the steps the model has taken, along with the output, within the Graphical User Interface (GUI). Developers using the model can also specify how long the model should consider a problem. Stats include:

• Context widow of 200,000 tokens
• Maximum output of 8,192 tokens
• The knowledge cutoff is November 2024
• Available on all Claude plans
• Achieved 79.3% accuracy on LegalBench benchmark testing

The Claude 3 and 3.5 model families include Opus (a high-performance model designed for handling complex analysis and lengthy tasks with multiple steps), Sonnet (which combines performance and speed for high-throughput tasks), and Haiku (a fast-responding model that executes lightweight actions with speed). Each is optimised for different use cases. These models are known for their nuanced understanding of context and ability to manage complex instructions. The Claude models can process text, images, and file uploads.

Claude 3.5 Sonnet

This model offers a strong balance of performance and cost-effectiveness. The introduction of this model also included ‘artifacts’ in the GUI. When a user prompts Claude for content, a dedicated window appears alongside the conversation, displaying the content output. Stats include:

• Context widow of 200,000 tokens
• Maximum output of 8,192 tokens
• The knowledge cutoff is April 2024
• Achieved 78.8% accuracy on LegalBench benchmark testing
• Outperformed OpenAI’s GPT-4 in several legal reasoning benchmarks

Claude 3.5 Haiku

An advancement of Haiku 3 with a fast-responding model for near-instant responsiveness. Stats include:

• The knowledge cutoff is July 2024
• Achieved 73.0% accuracy on LegalBench benchmark testing

Claude 3 Opus, Sonnet and Haiku

The Claude 3 family of models were released in March 2024. Three models were released in ascending order of capability: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus. The models allowed users to choose between balancing intelligence and speed. Stats include:

• Context widow of 200,000 tokens
• Maximum output of 7,096 tokens
• The knowledge cutoff is August 2023
• Opus was good for highly complex tasks and achieved 77.5% accuracy on LegalBench benchmark testing
• Sonnet strikes a balance between intelligence and speed, achieving an accuracy of 72.3% on LegalBench benchmark testing
• Haiku was the fastest model, featuring near-instant responses.