Claude Opus 4

O Anthropic

Założona 2021 · San Francisco, USA

Anthropic was founded in 2021 by Dario Amodei (CEO) and Daniela Amodei (President) after both left OpenAI together with several senior researchers including Tom Brown, Sam McCandlish and Jared Kaplan. The company is structured as a Public Benefit Corporation focused on AI safety research and the deployment of reliable, interpretable, and steerable AI systems. Anthropic's foundational research includes Constitutional AI (2022), Sparse Autoencoders for interpretability, and the influential paper 'Scaling Laws for Neural Language Models'. The Claude model family launched in March 2023, with major releases including Claude 2 (2023), Claude 3 family (Haiku, Sonnet, Opus in March 2024), Claude 3.5 Sonnet (June 2024), Claude 3.7 Sonnet with extended thinking (February 2025) and Claude 4 family (May 2025). Investors include Google ($3B+), Amazon (committed up to $8B), Spark Capital and Lightspeed, with total funding exceeding $15 billion and a 2025 valuation above $60 billion. Anthropic also publishes Responsible Scaling Policies and is a leading voice in ASL safety levels.

Architektura

Decoder-only Transformer (frontier reasoning model)

Claude Opus 4 is Anthropic's flagship reasoning model released in May 2025, optimized for agentic coding, multi-hour autonomous tasks and deep research. It uses a decoder-only Transformer architecture trained on a diverse multilingual corpus including curated web data, code repositories, books and licensed datasets, with an explicit data cutoff of March 2025. The model is post-trained with a combination of Reinforcement Learning from Human Feedback (RLHF), Constitutional AI (CAI) and Reinforcement Learning from AI Feedback (RLAIF), where a constitution of principles guides self-critique and revision. Opus 4 supports two operating modes: standard fast responses and extended thinking mode in which the model produces an internal reasoning trace before its final answer. Tool use, parallel tool calling and long-running agent loops are first-class behaviours, with native support for Anthropic's computer use API (screen control, mouse, keyboard). Vision grounding was trained jointly with text so that images, charts and PDFs can be reasoned about within the same context window. The model was evaluated against ASL-3 safeguards under Anthropic's Responsible Scaling Policy and shipped with classifier-based safety filters covering CBRN, cyberweapons, and child safety.

Parametry

Undisclosed (estimated multi-hundred billion parameters dense)

Kontekst

200 000 tokenów

Możliwości

• State-of-the-art coding performance on SWE-bench Verified (~72% on agentic eval at launch)
• Extended thinking mode for chain-of-thought reasoning on hard math and science problems
• Sustained multi-hour autonomous coding sessions with tool use and self-correction
• Native computer use: screenshots, mouse and keyboard control of a virtual desktop
• 200K token context window with strong needle-in-haystack recall across the full window
• Multimodal input: text, images, PDFs, charts and diagrams in a single prompt
• Parallel tool calling and stable multi-turn agent loops
• Constitutional-AI alignment with low refusal rate on benign questions
• Strong instruction following including XML tags, schemas and strict JSON output
• Multilingual fluency across English, German, French, Spanish, Japanese, Chinese and more
• Best for: complex coding agents, research synthesis from large document sets, regulated-industry copilots, autonomous workflows.

Trening i licencja

Pretrained on a curated mix of publicly available internet text, licensed third-party data, code from public repositories and human-generated conversations. Anthropic does not disclose exact token counts but the corpus is multi-trillion tokens with heavy filtering for quality, deduplication and copyright. The data cutoff is March 2025. Post-training uses RLHF, Constitutional AI and RLAIF on a much smaller, high-quality set of human-rated comparisons and synthetic constitutional critiques.

Licencja: Proprietary commercial license via Anthropic API, Amazon Bedrock and Google Vertex AI. Commercial use permitted under Anthropic's Usage Policy.

Testy bezpieczeństwa: Deployed under Responsible Scaling Policy ASL-3 safeguards with constitutional classifiers, automated red-teaming, third-party evaluations by METR and Apollo Research, and a Bug Bounty programme.