Anthropic announced Claude 3.5 Sonnet on June 21, 2024. This is the first release in the Claude 3.5 model family and, according to the company, raises performance expectations for language and vision tasks while maintaining mid-tier speed and cost characteristics.
Availability and pricing
Claude 3.5 Sonnet is available for free on Claude.ai and in the Claude iOS app; Claude Pro and Team subscribers receive significantly higher rate limits. The model is also offered via the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI. Anthropic’s stated pricing is $3 per million input tokens and $15 per million output tokens. The model supports a 200,000-token context window.
Performance: reasoning, coding, and speed
Anthropic says Claude 3.5 Sonnet sets new benchmarks on multiple fronts, including graduate-level reasoning (GPQA), undergraduate-level knowledge (MMLU), and coding performance (HumanEval). The company highlights improved handling of nuance, humor, and complex instructions, and stronger capability to produce natural, high-quality prose.
Claude 3.5 Sonnet runs at twice the speed of Claude 3 Opus, according to Anthropic. That speed improvement combined with cost-effective pricing makes the model suitable for complex, context-sensitive applications such as customer support and orchestrating multi-step workflows.
In an internal agentic coding evaluation, Claude 3.5 Sonnet solved 64% of problems compared with 38% for Claude 3 Opus. The evaluation measured the model’s ability to fix bugs or add functionality to an open-source codebase based on a natural language description. When provided with appropriate instructions and tools, Claude 3.5 Sonnet can write, edit, and execute code autonomously and perform sophisticated reasoning and troubleshooting—useful for updating legacy applications and migrating codebases.
State-of-the-art vision
Anthropic describes Claude 3.5 Sonnet as its strongest vision model to date, outperforming Claude 3 Opus on standard vision benchmarks. The largest gains appear in tasks that require visual reasoning, such as interpreting charts and graphs. The model also more accurately transcribes text from imperfect images, a capability relevant for retail, logistics and financial services where insights may come from images, graphics or illustrations rather than text alone.
Artifacts: a new workspace on Claude.ai
Anthropic introduced a feature called Artifacts on Claude.ai alongside Sonnet. When users ask Claude to generate content—code snippets, text documents, or website designs—those outputs appear in an Artifacts pane next to the conversation. This creates a dynamic workspace where users can view, edit and build on Claude’s outputs in real time, making it easier to integrate AI-generated content into projects and workflows.
The feature marks a shift from a purely conversational AI toward a collaborative work environment. Anthropic says it will extend Claude.ai to support team collaboration: in the near term teams, and later entire organizations, will be able to centralize knowledge, documents and ongoing work in a shared space with Claude functioning as an on-demand teammate.
Safety and privacy commitments
Anthropic states that its models undergo rigorous testing and are trained to reduce misuse. Despite the intelligence gains in Claude 3.5 Sonnet, internal red teaming assessments place the model at ASL-2; further details are available in the model card addendum.
The company engaged external experts to test and refine safety mechanisms. Anthropic provided Claude 3.5 Sonnet to the UK’s Artificial Intelligence Safety Institute (UK AISI) for pre-deployment safety evaluation; the UK AISI completed tests and shared results with the US AI Safety Institute (US AISI) under a memorandum of understanding between the two institutes.
Anthropic also incorporated policy feedback from outside subject-matter experts to improve evaluations against new patterns of misuse. For example, feedback from child safety experts at Thorn was used to update classifiers and fine-tune models. The company reiterates that it does not train its generative models on user-submitted data unless a user gives explicit permission.
What’s next
Anthropic says it aims to materially improve the trade-off between intelligence, speed and cost every few months. To complete the Claude 3.5 family, it plans to release Claude 3.5 Haiku and Claude 3.5 Opus later this year.
The company is also developing new modalities and enterprise integrations, and exploring features such as Memory, which would let Claude retain user preferences and interaction history as specified to provide a more personalized and efficient experience.
Anthropic encourages users to submit feedback on Claude 3.5 Sonnet in-product to inform the development roadmap and help improve the model. The company said it looks forward to seeing how people use Claude in their projects and workflows.
