Claude Sonnet 4

Acerca de Anthropic

Fundado en 2021 · San Francisco, USA

Anthropic was founded in 2021 by Dario and Daniela Amodei alongside ex-OpenAI researchers including Tom Brown and Jared Kaplan. The company is a Public Benefit Corporation focused on AI safety, interpretability and steerability. Foundational papers from the team include 'Scaling Laws for Neural Language Models' (Kaplan et al., 2020), 'Training a Helpful and Harmless Assistant' (2022) and 'Constitutional AI' (2022). The Claude product family was launched in March 2023 starting with Claude 1, followed by Claude 2, the Claude 3 family in March 2024, Claude 3.5 Sonnet in mid-2024, Claude 3.7 Sonnet with extended thinking in early 2025 and the Claude 4 family (Opus 4 and Sonnet 4) in May 2025. Anthropic is backed by Google, Amazon, Spark Capital, Lightspeed and others, with cumulative funding above $15 billion and a 2025 valuation above $60 billion. Anthropic publishes a Responsible Scaling Policy that defines ASL safety levels and is widely cited in AI governance debates.

Arquitectura

Decoder-only Transformer (balanced production model)

Claude Sonnet 4 is the mid-tier flagship in the Claude 4 family, positioned as the default daily-driver model balancing latency, cost and capability. Architecturally it is a decoder-only Transformer pretrained on a curated multilingual corpus of web data, source code, books and licensed datasets, with a March 2025 data cutoff. Post-training applies a combination of supervised fine-tuning, Reinforcement Learning from Human Feedback (RLHF), Constitutional AI for harmlessness, and Reinforcement Learning from AI Feedback to scale alignment work. Sonnet 4 supports extended thinking with adjustable thinking budgets, allowing developers to trade additional latency for stronger reasoning on demand. Tool use, parallel tool calls, structured JSON output and Anthropic's computer use API are supported natively. Multimodal training jointly grounds images and text so charts, screenshots and PDF pages can be reasoned about within the 200K context. Sonnet 4 is co-trained alongside Opus 4 using shared infrastructure, then distilled and specialised toward throughput rather than maximum capability. Anthropic evaluated the model under Responsible Scaling Policy ASL-3 safeguards with internal red-teaming and external partners such as METR and Apollo Research before public release.

Parámetros

Undisclosed (estimated mid-hundred billion parameters dense)

Contexto

200.000 tokens

Capacidades

• Strong coding model with high SWE-bench Verified scores at production cost
• Optional extended thinking mode with configurable token budget
• 200K context window with reliable long-document recall
• Multimodal input including images, charts and PDFs
• Tool use, parallel tool calls and computer use API
• Stable agent loops over hours of autonomous work
• Strict JSON, XML and schema-constrained output
• Excellent at instruction following and persona retention
• Multilingual reasoning across major European and Asian languages
• Low refusal rate on benign technical content thanks to Constitutional AI tuning
• Best for: production assistants, coding copilots, agent backbones, document analysis at scale.

Entrenamiento y licencia

Pretrained on a multi-trillion-token mixture of curated web text, books, code repositories, licensed data and synthetic data, with quality and deduplication filters. Data cutoff is March 2025. Post-training uses human preference data, Constitutional AI critiques and RLAIF on top of supervised fine-tuning.

Licencia: Proprietary commercial license via Anthropic API, Amazon Bedrock and Google Vertex AI.

Pruebas de seguridad: Anthropic Frontier Red Team plus external evaluators (METR, Apollo Research) tested for autonomy, CBRN uplift and cyber misuse under ASL-3 safeguards.