Anthropic Announces Claude 3 Model Family: Opus, Sonnet, and Haiku

Anthropic announced the release of the Claude 3 model family, which sets new industry benchmarks across a wide range of cognitive tasks. The family consists of three progressively capable models: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus, each providing a different trade‑off between intelligence, speed, and cost.

Opus and Sonnet are already available on claude.ai and through the Claude API, which is now generally accessible in 159 countries. Haiku will be launched soon.

Intelligent new standard – Opus, the most capable model, surpasses peers on benchmarks such as MMLU (undergraduate‑level knowledge), GPQA (graduate‑level reasoning), and GSM8K (basic mathematics), demonstrating near‑human understanding and fluency on complex tasks.

All Claude 3 models show stronger performance in analysis and prediction, nuanced content creation, code generation, and multilingual dialogue (including Spanish, Japanese, and French).

Near‑real‑time results – The models support real‑time customer chat, auto‑completion, and data extraction tasks that require instantaneous responses. Haiku is the fastest and most cost‑effective, capable of reading a 10k‑token arXiv paper with figures in under three seconds.

Sonnet runs about twice as fast as Claude 2/2.1 while delivering higher intelligence, making it ideal for rapid‑response workloads such as knowledge retrieval or sales automation. Opus matches the speed of Claude 2/2.1 but with higher intelligence.

Powerful visual abilities – Claude 3 models handle diverse visual formats (photos, charts, diagrams, technical schematics) and are suited for enterprises where up to 50 % of knowledge resides in PDFs, flowcharts, or slide decks.

Fewer refusals – Compared with earlier Claude models, Opus, Sonnet, and Haiku significantly reduce unnecessary refusals near system safety boundaries, showing finer understanding of harmless prompts.

Improved accuracy – In challenging factual question sets, Opus achieves roughly double the correct‑answer rate of Claude 2.1 while cutting hallucinations. Future releases will add citation capabilities to point to exact source sentences.

Long context and near‑perfect memory – The family ships with a 200 K token context window, and all three models can accept over one million tokens for selective customers. The “Needle‑in‑a‑Haystack” (NIAH) benchmark shows Opus attaining >99 % recall accuracy.

Responsible design – Dedicated teams monitor risks such as misinformation, CSAM, bio‑misuse, election interference, and autonomous replication. Techniques like Constitutional AI and ongoing bias mitigation (e.g., BBQ benchmark) reduce partisan bias.

The models are classified at AI Safety Level 2 (ASL‑2) under Anthropic’s Responsible Scaling Policy, with red‑team assessments indicating negligible catastrophic risk. Ongoing monitoring aims toward ASL‑3.

Ease of use – Claude 3 models follow complex, multi‑step instructions better, adhere to brand voice, and generate structured outputs like JSON, simplifying tasks such as natural‑language classification and sentiment analysis.

Model details – Claude 3 Opus is the most intelligent, excelling on highly complex tasks. Claude 3 Sonnet balances intelligence and speed for enterprise workloads at lower cost. Claude 3 Haiku is the fastest, compact model for near‑instant responses.

Availability – Opus and Sonnet are available via the API (global rollout) and to Claude Pro subscribers; Haiku will follow. Sonnet is also on Amazon Bedrock and in private preview on Google Cloud Vertex AI Model Garden, with Opus and Haiku slated for both platforms.

Anthropic plans frequent updates to the Claude 3 family, adding features such as tool use (function calling), interactive coding (REPL), and advanced agent capabilities, while continuing to strengthen safety measures.

For more information and to start building with Claude, visit anthropic.com/claude .