Claude Sonnet 4.6 Unveiled: Enhanced Coding and Computer Use Capabilities

Anthropic has introduced Claude Sonnet 4.6, the latest iteration of its Sonnet model line, showcasing significant enhancements in coding, computer use, long-context reasoning, and more. Key among its new features is a 1M token context window, currently in beta, which extends its ability to handle substantial data volumes such as entire codebases and lengthy documents.
Sonnet 4.6 is now the default model in both claude.ai and Claude Cowork for users subscribed to Free and Pro plans, with pricing starting at $3/$15 per million tokens, unchanged from Sonnet 4.5. Early access developers have noted marked improvements in instruction following, consistency, and task execution, finding Sonnet 4.6 preferable over the November 2025's Claude Opus 4.5 for many real-world tasks.
In terms of computer use skills, Sonnet 4.6 builds on the progress made since Claude's first general-purpose computer-using model release in October 2024. The model interacts with software as a human does, using virtual interfaces like a mouse and keyboard without specialized connectors. OSWorld benchmarks, a standard for AI computer use, confirm its development, showing near-human capabilities in complex tasks like navigating spreadsheets and completing web forms.
Security remains a focus, with Sonnet 4.6 undergoing extensive safety evaluations indicating its improved resilience to prompt injection attacks, rivaling the Opus-class models in this regard. The system card and API documentation provide more details and guidance on these safety enhancements.
📖 Read the full source: HN AI Agents
👀 See Also

70% of devs say AI code has more vulns; 30% ship it anyway — Checkmarx survey
70% of developers believe AI-generated code has significantly more vulnerabilities, yet 30% knowingly ship vulnerable code into production. The Checkmarx survey of 2,350 respondents also finds 93% of orgs suffered security breaches from vulnerable apps.

Melbourne Psychiatrist Refuses New Patients Who Don't Consent to AI Note-Taking
A Melbourne psychiatrist now requires new patients to consent to AI transcription for sessions or be referred elsewhere, raising data security and accuracy concerns.

GPT-5.5 on OpenClaw: Users Report 5-Hour Limit Burned in 3 Prompts, Broken Code
Reddit user reports GPT-5.5 on OpenClaw consumed 5-hour limit in 3 prompts, introduced bugs; DeepSeek V4 Pro fixed the mess.

Weekly Multimodal AI Roundup: Holotron-12B, Nemotron Omni, GlyphPrinter, and More
This week's multimodal AI highlights include Holotron-12B for computer-use tasks, NVIDIA's Nemotron Omni models integrating language+vision+voice, GlyphPrinter for accurate text rendering in image generation, and several open-source projects for video enhancement, 3D segmentation, and multi-agent systems.