Claude Sonnet 5 Is HERE – Hands-On With Anthropic’s NEW Model!

Summary
In this hands-on evaluation, presenter Bijan Bowen reviews Anthropic’s Claude Sonnet 5 alongside the Claude desktop app beta for Linux. Running benchmarks and interactive coding tests via Claude Code and the Claude web interface, Bowen examines how Sonnet 5 performs on complex 3D web applications, games, and agentic tasks compared to prior Opus and Sonnet models.

What is shown

• Anthropic Announcement & Pricing [00:11–02:14]: Overview of Anthropic's blog post "Introducing Claude Sonnet 5" (dated June 30, 2026), reviewing benchmark tables, new tokenizer details, and pricing structure ($2/$10 introductory per million tokens through August 31, 2026, then $3/$15).
• Effort Levels & Web UI [03:40–04:19]: Demonstrating effort level settings in the Claude web interface, showing Sonnet 5's new "Max" effort option alongside existing tiers (Low, Medium, High, Extra).
• BrowserOS Benchmark [04:40–10:49]: A single-file web OS generated via Claude Code featuring file management, a terminal with built-in commands (Matrix effect, jokes), a paint application, a functional procedural music player, and two 3D games (Crime City 3D and Zombie Siege 3D), plus a voice-controlled "Echo Assistant."
• 3D Skateboarding Game [10:50–12:40]: Testing a C++ 3D skateboarding game (Cali Skate) built with Claude Code on "Ultracode" setting in 19 minutes, 25 seconds, showing tricks (kickflips, heelflips, shovits), NPC pedestrians, and environment physics.
• 3D Subway Station & FPS Conversion [12:41–15:30]: Generating a detailed 3D subway station scene (Maplewood Jct.) on Max effort in Three.js, followed by converting it into a playable first-person shooter (Last Stop) with zombie enemies, sound effects, and weapon mechanics.
• 3D Skydiving Simulator [15:31–18:58]: Evaluating Dropzon, a skydiving game featuring freefall physics, variable wind sound effects, and an automatic parachute deployment feature.
• Interactive 3D Watch Brand Site [19:03–21:02]: Generating a promotional website for fictional watchmaker Slappis, including a procedural 3D watch model with pan animations in the hero header.
• Time-Traveling 3D City Block [21:03–25:23]: A 3D urban environment featuring a slider transitioning between historical eras (1945, 1965, 1985 synthwave aesthetic, 2005, and 2025).
• 3D Laptop Model from Photos [25:24–27:32]: Attempting a multimodal task to replicate a custom 3D-printed laptop from a folder of reference photographs.
• F1 Racing Game & 3D Drum Kit [27:33–30:36]: Testing an F1 racer (Apex Circuit) on Medium effort, and an interactive 3D drum kit (Studio Kit) on High effort featuring interactive pads and automated rhythm playback presets.
• Terrain Driving Simulator [30:37–31:56]: Generating Ridgeback, an off-road driving simulation reusing terrain generation code originally created by Claude Opus 4.8.

Claims & numbers

• The presenter notes Claude Sonnet 5's introductory pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which it increases to standard pricing of $3 per million input and $15 per million output tokens [01:36].
• Anthropic's footnotes indicate Sonnet 5 uses an updated tokenizer where text maps to roughly 1.0–1.35x more tokens depending on content type [02:03].
• The presenter cites official benchmark comparisons: SWE-bench Pro scores are 63.2% for Sonnet 5, 58.1% for Sonnet 4.6, and 69.2% for Opus 4.8 [01:03]; Terminal Bench 2.1 agentic coding scores are 80.4% for Sonnet 5, 67.0% for Sonnet 4.6, and 82.7% for Opus 4.8 [02:37].
• Opus 4.7 benchmarks shown on Anthropic's announcement page scored 64.3% on SWE-bench Pro and 69.4% on Terminal Bench 2.1 [02:30].
• The C++ skateboarding game task took 19 minutes and 25 seconds across 18 agent tasks and 355.4k tokens using Claude Code [10:50].

Notable quotes

• "What happens if we just don't deploy the parachute? So... oh, it automatically deploys for us. What an Anthropic thing to do!" [18:35]
• "Overall, I have to say, honestly, I'm not impressed. I don't know what I was expecting. It is a Sonnet-class model which is in the middle tier of intelligence of the publicly available Anthropic models..." [32:05]
• "It's just so incredibly slow to use on any decently capable thinking level, which was kind of a letdown." [32:28]

Assessment
This is an independent community hands-on review and stress-test of Claude Sonnet 5 across various agentic coding and 3D rendering prompts. The testing is conducted live in real time using local and web interfaces, showing genuine flaws, execution delays, and model rendering bugs alongside functional elements.

Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.