Claude Opus 4.7

Sobre Anthropic

Fundado em 2021 · San Francisco, USA

Anthropic was founded in 2021 by Dario Amodei (CEO) and Daniela Amodei (President) together with senior researchers Tom Brown, Sam McCandlish, Jared Kaplan and others who left OpenAI. The company is structured as a Public Benefit Corporation focused on AI safety and the deployment of reliable, interpretable, steerable AI systems. Anthropic's foundational research includes Constitutional AI (2022), Sparse Autoencoders for interpretability, and influential scaling laws work. The Claude model family launched in March 2023, followed by Claude 2 (2023), Claude 3 (March 2024), Claude 3.5 Sonnet (June 2024), Claude 3.7 Sonnet with extended thinking (February 2025), the Claude 4 family (May 2025), Claude 4.5 (late 2025), Claude Sonnet 4.6 (February 2026) and Claude Opus 4.7 (April 2026). Investors include Google ($3B+), Amazon (committed up to $8B), Spark Capital and Lightspeed; total funding exceeds $20 billion with a 2026 valuation above $150 billion. Anthropic publishes Responsible Scaling Policies and is a leading voice on ASL safety levels.

Arquitetura

Decoder-only Transformer (frontier reasoning model with self-verification)

Claude Opus 4.7 is Anthropic's flagship frontier model released April 16, 2026. It builds on the Opus 4 lineage with three headline upgrades: native output self-verification (the model spends test-time compute to check its own answers before returning them), 3x higher vision input resolution, and an upgraded tokenizer (which can produce up to 35% more tokens for the same input text). Architecturally it remains a decoder-only Transformer optimized for agentic engineering and long-horizon tasks. Pretraining used a multi-trillion-token mixture of curated web text, code repositories, books, scientific papers and licensed data with a knowledge cutoff in early 2026. Post-training combined RLHF, Constitutional AI (CAI), RLAIF, and large-scale reinforcement learning against verifiable rewards on coding, math and tool-use trajectories. The model retains the two operating modes from Opus 4: standard fast responses and extended thinking mode with adjustable thinking budget. Native computer use (screenshots, mouse and keyboard control), parallel tool calling, and stable multi-hour autonomous agent loops are first-class behaviours. Vision input was jointly trained with text and now supports denser document, diagram and screen reading. The model shipped under Anthropic's Responsible Scaling Policy ASL-3 safeguards with constitutional classifiers, external evaluations and a public Bug Bounty programme.

Parâmetros

Undisclosed (estimated multi-hundred billion parameters dense, hybrid sparse layers)

Contexto

200.000 tokens

Capacidades

• 87.6% on SWE-bench Verified at launch, top of the agentic coding leaderboard
• Native output self-verification: the model checks its own answers before returning
• 3x higher vision input resolution for dense documents and screenshots
• Extended thinking mode with adjustable thinking budget for hard reasoning
• Sustained multi-hour autonomous coding sessions with stable tool-use loops
• Native computer use: screenshots, mouse and keyboard control of a virtual desktop
• 200K token context window with strong needle-in-haystack recall
• Multimodal input: text, images, PDFs, charts and diagrams in a single prompt
• Parallel tool calling and structured JSON / XML output
• Constitutional-AI alignment with low refusal rate on benign questions
• Multilingual fluency across English, German, French, Spanish, Japanese, Chinese and more
• Best for: complex coding agents, regulated-industry copilots, autonomous workflows, long-document analysis.

Treinamento & licença

Pretrained on a curated multi-trillion-token mixture of publicly available internet text, licensed third-party data, code from public repositories and human-generated conversations. Anthropic does not disclose exact token counts; the corpus is heavily filtered for quality, deduplication and copyright. Knowledge cutoff is early 2026. Post-training uses RLHF, Constitutional AI, RLAIF, and reinforcement learning against verifiable rewards on coding, math and agent trajectories.

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

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