On March 4, 2024, Anthropic unveiled the Claude 3 family of models, comprising three variants: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus. The lineup is designed so users can choose between different trade-offs of intelligence, latency and cost depending on the application.
Availability and deployment
Claude 3 Opus and Claude 3 Sonnet are available now via claude.ai and the Claude API, which Anthropic says is generally available in 159 countries. Claude 3 Haiku is expected to be released soon. Sonnet powers the free experience on claude.ai, while Opus is offered to Claude Pro subscribers. Sonnet is also available through Amazon Bedrock and in private preview on Google Cloud’s Vertex AI Model Garden; Opus and Haiku will be added to those services later.
Model distinctions
-
Claude 3 Opus: positioned as the most capable model, Anthropic reports it outperforms peers on many standard benchmarks, including MMLU (undergraduate-level knowledge), GPQA (graduate-level reasoning) and GSM8K (basic math). Intended use cases include task automation across APIs and databases, R&D and strategy work. Pricing listed: input $15 per million tokens, output $75 per million tokens. Context window: 200K tokens (with 1M tokens available in certain cases).
-
Claude 3 Sonnet: intended to balance intelligence and speed for enterprise workloads, optimized for large-scale deployments with better cost efficiency. Pricing: input $3 per million tokens, output $15 per million tokens. Context window: 200K tokens. Suggested uses include retrieval-augmented generation (RAG), sales support and time-saving tasks such as code generation.
-
Claude 3 Haiku: the smallest, fastest model designed for near-instant responses and low-cost interactions. Pricing: input $0.25 per million tokens, output $1.25 per million tokens. Context window: 200K tokens. Targeted at live customer support, quick translations, moderation and cost-saving operations.
Latency and accuracy claims
Anthropic emphasizes near-instant results for many real-time tasks. They claim Haiku is the fastest and most cost-effective model in its intelligence tier and can process an information-dense arXiv paper of roughly 10k tokens with charts and graphs in under three seconds. Sonnet is reported to be roughly 2× faster than Claude 2 and Claude 2.1 on the majority of workloads while offering higher intelligence; Opus achieves similar speeds to Claude 2 family models with substantially higher capability.
The company also reports improved accuracy: on a set of complex factual questions designed to probe known model weaknesses, Opus achieved roughly twice the number of correct answers compared to Claude 2.1 and reduced the rate of incorrect answers (hallucinations). Anthropic says it will soon enable citations in Claude 3 so responses can point to precise sentences in reference material.
Long context windows and recall performance
At launch, Claude 3 models offer a 200K token context window. Anthropic states that all three models are capable of accepting inputs exceeding one million tokens and that expanded capacity could be made available to select customers. On the 'Needle In A Haystack' (NIAH) recall evaluation, Claude 3 Opus reportedly achieved near-perfect recall (over 99% accuracy) and in some instances identified artifacts in the evaluation data, such as sentences that appeared artificially inserted.
Vision and multimodal processing
Claude 3 models include vision capabilities able to handle photos, charts, graphs and technical diagrams. Anthropic highlights this modality for enterprise customers that store a significant portion of knowledge in formats like PDFs, flowcharts or slides.
Refusals, bias and safety work
According to Anthropic, the Claude 3 models refuse fewer prompts unnecessarily than prior generations and display a more nuanced understanding of requests—rejecting content that poses real harm while answering harmless borderline prompts more often. The company reports reduced bias on the Bias Benchmark for Question Answering (BBQ) compared with prior Claude versions.
Anthropic describes ongoing risk mitigation across misinformation, CSAM, biological misuse, election interference and autonomous replication. The Claude 3 family is classified at AI Safety Level 2 (ASL-2) under Anthropic’s Responsible Scaling Policy. Their red teaming evaluations, carried out in line with the company’s White House commitments and the 2023 US Executive Order, concluded the models currently present negligible potential for catastrophic risk; Anthropic will continue monitoring future models for proximity to an ASL-3 threshold. Further safety details are provided in the Claude 3 model card.
Usability and upcoming features
The company says Claude 3 models better follow complex, multi-step instructions, adhere to brand voice and produce structured outputs like JSON more reliably. Anthropic plans to roll out additional features over the coming months targeted at enterprise use and large-scale deployments, including Tool Use (function calling), interactive coding (REPL) and more advanced agentic capabilities.
Summary
The Claude 3 family aims to serve a range of use cases: Haiku for low-latency, low-cost interactions; Sonnet for scalable enterprise workloads; and Opus for highest-accuracy, high-complexity tasks. Anthropic highlights extended context handling, faster response times, improved factual accuracy and continued investment in safety and bias mitigation as key elements of the release.
