Claude Sonnet 4.6

O Anthropic

Założona 2021 · San Francisco, USA

Anthropic was founded in 2021 by Dario and Daniela Amodei together with senior researchers from OpenAI. The company is a Public Benefit Corporation focused on AI safety. Key research milestones include Constitutional AI (2022), Sparse Autoencoders for interpretability, and the Claude model family launched in March 2023. Claude Sonnet has been Anthropic's flagship 'balanced' tier since the Claude 3 family, with major releases including Sonnet 3.5 (June 2024), Sonnet 3.7 with extended thinking (February 2025), Sonnet 4 (May 2025), Sonnet 4.5 (late 2025) and Sonnet 4.6 (February 17, 2026). Investors include Google, Amazon, Spark Capital and Lightspeed; total funding exceeds $20 billion with a 2026 valuation above $150 billion.

Architektura

Decoder-only Transformer (balanced production model with extended thinking)

Claude Sonnet 4.6 was released February 17, 2026 as the mid-tier workhorse of the Claude 4.x family. It is a decoder-only Transformer engineered to deliver near-Opus quality at one-fifth the cost. Pretraining used the same multi-trillion-token mixture as Opus 4.7 with a knowledge cutoff in early 2026. Post-training combined RLHF, Constitutional AI, RLAIF, and reinforcement learning against verifiable rewards on coding, math and tool-use trajectories. Sonnet 4.6 supports the full Opus feature set: extended thinking mode, native vision, computer use, parallel tool calling, and 200K context. The training emphasis was on efficiency: Sonnet 4.6 outperforms Opus 4 on most coding and writing benchmarks while running ~2x faster and at one-fifth the cost. Anthropic positions Sonnet 4.6 as the recommended default for production workloads.

Parametry

Undisclosed (estimated dense model in the 100B-range)

Kontekst

200 000 tokenów

Możliwości

• Near-Opus quality on most coding, writing and analysis tasks at 5x lower cost
• Extended thinking mode with adjustable thinking budget
• 200K token context window with strong needle-in-haystack recall
• Multimodal input: text, images, PDFs, charts and diagrams
• Native computer use: screenshots, mouse and keyboard control
• Parallel tool calling and stable multi-turn agent loops
• Strong instruction following including XML tags, schemas and strict JSON
• Constitutional-AI alignment with low refusal rate on benign questions
• Multilingual fluency across 20+ languages
• ~2x faster latency than Opus 4.7 on equivalent tasks
• Best for: production agents, developer copilots, customer support, most enterprise API scenarios.

Trening i licencja

Pretrained on a curated multi-trillion-token mixture identical to the Opus 4.7 corpus, with heavy quality filtering, deduplication and copyright screening. Knowledge cutoff in early 2026. Post-training uses RLHF, Constitutional AI, RLAIF and RL against verifiable rewards.

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 and a public Bug Bounty programme.