Claude Sonnet 4.6: Near Opus-Level Intelligence at Half the Price

WorldofAI puts Claude Sonnet 4.6 through a comprehensive set of tests, from front-end generation to Minecraft clones to browser automation.

Key Specs & Pricing

• Full upgrade across coding, computer use, long context reasoning, agent planning, knowledge work, and design
• 1 million token context window (beta)
• Same pricing as Sonnet 4.5: $3/M input tokens**, **$6/M output tokens — half the price of Opus 4.6
• Scored 79.6 on SWE-Bench Verified, achieving state-of-the-art on agentic coding and financial analysis benchmarks
• Around 2x faster than Opus
• Less hallucination, better instruction following, and improved long context handling

How to Access

• API — available directly
• Claude chatbot — free but heavily rate-limited
• LM Arena — free access
• Kilo Code — open-source agent, provides $25 free credits
• Open Router — API access

Front-End Generation

• Generated a SaaS landing page with a new visual style not seen in other models
• Strong typography, color palette, and component structure
• Outperformed Gemini’s front-end output in the reviewer’s opinion

Mac OS Clone

• Single-prompt generation of a functional Mac OS-style interface
• Working components: Finder, file manager, terminal, calculator, system settings
• Functional features: music player with time bar, spotlight search, Wi-Fi/battery/brightness controls, wallpaper changing, folder management
• Top bar tools mimicking real Mac OS behavior

Minecraft Clone (Kilo Code + Free API)

• Deployed multiple agents via Kilo Code to build “Boxelcraft”
• First generation with actual terrain generation (including underground caves)
• Features: heart bar, food bar, block placement/breaking, world creation, configurable settings
• Notably faster than Opus at code generation
• Some bugs with block formatting and lag in browser — improved after re-prompting for lower specs

3D Simulations

• F1 Car Drifting: 3D simulation with drift marks, smoke from rear tires only, RPM/speed display, multiple camera angles (aerial, sidecam). Cleaner and more intuitive than Opus output
• 3D Room Design: Orbit control, furniture toggle on/off, night mode visualization. Decent output in Three.js
• Marble Labyrinth: Board-tilt game controlled by mouse movement, with physics and checkpoints. Fully functional

SVG Generation

• Tested butterfly, robot, pelican on a bike, PS5 controller
• Decent output but did not match Opus 4.6 quality for SVG specifically
• SVG remains a relative weakness compared to Opus

Browser Automation (Stress Test)

• Autonomous multi-task setup via Kilo Code: HTML dashboard, Python automation script, CSV data generation, task list
• Used Playwright to open a browser, Google search for latest AI news, and scrape top 5 headlines
• Connected Python script to the front-end dashboard displaying weather, CSV data, and a number grid
• Fast and reliable execution across all components

Conclusion

• Best bang for the buck — near Opus intelligence at half the cost
• Strong at agentic use cases, browser automation, and reasoning across large contexts
• Better instruction following than previous Sonnet models and Gemini
• 1 million context window is a significant advantage for complex tasks
• SVG generation still trails Opus 4.6