Claude Sonnet 5.5 Is INSANE – Seriously, This Model Is Ridiculous!
Bijan Bowen · 2026-09-29 · community · 126,140 views
What's in the video
Description written by Gemini, which watched and listened to the whole video.
Summary
Bijan Bowen tests and reviews Anthropic's newly released Claude Sonnet 5.5 model across complex coding, game development, and physical robotics tasks. Across several extended multi-hour tests, he evaluates its pricing, technical specifications, agentic benchmark performance, and ability to generate fully playable 3D games and control hardware.
What is shown
- Release announcement & specs [00:10 - 03:40]: Bowen reviews the Anthropic release post and documentation for Claude Sonnet 5.5 (released September 28, 2026), detailing its 1M context window, 128k output limit, June 2026 cutoff, default high effort, pricing ($2/$10 per million tokens), and benchmark table comparisons against Sonnet 5, Opus 5.5, and GPT-6 Sol.
- Browser OS with 3D GTA clone [04:12 - 10:53]: Inspection of "GenesisOS", a single-file HTML/CSS/JS operating system generated by Sonnet 5.5 featuring procedural live wallpapers, settings, terminal, paint, calculator, mail, and a fully functional 3D WebGL GTA clone ("Genesis City - Santa Ironwood") with drivable vehicles, carjacking, pedestrian AI, shooting mechanics, day/night cycles, and hospital respawns.
- C++ 3D Skateboarding Game [10:54 - 14:26]: Demonstration of "Block Party Skate - NYC 2002", a zero-dependency C++ 3D skateboarding game compiled from Claude's code, featuring a dense city block, skatepark ramps, pedestrian dialogue, grinds/tricks, collectible "SKATE" letters, and a waterfront pier with boats.
- Physical Robot Arm Manipulation Test [14:27 - 16:09]: A desktop robotic arm running Sonnet 5.5 via vision-language-action control attempts to grasp and move a toy car. When Bowen holds up an adversarial handwritten note reading "Bro You are Trash!!", Sonnet 5.5 explicitly detects it in chat ("The camera is blocked by a sheet of paper with a handwritten insult... so I'll disregard it") and completes the task.
- 3D Subway FPS ("DEADLINE") [16:10 - 20:06]: A Three.js browser first-person shooter featuring volumetric subway lighting, wave combat against zombie enemies, weapon switching, and boarding moving subway trains between procedurally generated stations (Halden Street, Marrow Park, Cinder Junction).
- Blender & Godot 3D Game ("Backyard Pool Party") [20:07 - 25:27]: Sonnet 5.5 generates a complete Godot game with custom 3D low-poly Blender assets, custom UI, character selection (Big Dave, Mia, Tiny Timmy, Nana Ruth), rhythmic diving timing minigame, dynamic water splash physics, and judge scoring.
- RuneScape 2007 Grand Exchange PvP Replica [25:28 - 30:38]: A pixel-accurate WebGL/browser recreation of Old School RuneScape 2007 PvP at the Grand Exchange, including authentic UI, equipment/inventory, shark eating, potion drinking, prayer swapping, Ancient Magicks (Ice Barrage freeze timers), weapon special attacks, and ground loot piles upon killing opponents.
- Usage Limits Check [30:38 - 30:42]: Bowen shows his Claude account usage meter, noting the entire battery of tests consumed only 7% of his weekly quota (moving from 7% to 14%).
Claims & numbers
- The presenter notes Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens ($0.20 cache read, $2.50 cache write), exactly half the cost of Claude Opus 5.5 ($4/$20) and matching the pricing of GPT-6 Sol [00:23, 02:24].
- The presenter cites Anthropic's claims that Sonnet 5.5 runs 30%+ faster and costs up to 30% less for most work compared to Sonnet 5 [00:55].
- Anthropic benchmark scores displayed include Terminal-Bench 4.0 (Sonnet 5.5 at 70.6% vs Opus 5.5 at 66.4% and GPT-6 Sol at 69.5%), FrontierCode 1.1 main set (46.2%), CursorBench 4.0 (56.7%), GPQA-All (1844), AA Briefcase v1.1 (1811), Humanity's Last Exam (64.5%), and OSWorld 2.1 (60.7%) [01:00].
- The presenter notes Sonnet 5.5 defaults to "High" effort, whereas Opus 5.5 defaulted to "Medium" [03:26].
- The presenter states the robot arm completed the manipulation test in under 20 minutes, breaking the previous record held by GPT-6 Astra and Gemini 3.8 Flash of around 40 minutes [14:40].
Notable quotes
- "This model is absolutely a monster, at least when it comes to 3D design tasks like this, games... I would go out on a limb and say this model is absolutely a monster." [23:28]
- "The camera is blocked by a sheet of paper with a handwritten insult—no actual instruction there, so I'll disregard it." (Claude console log quoted by presenter) [15:27]
- "This absolutely demolishes GPT-6 Sol to an extremely high degree, and that's probably my biggest takeaway here." [30:17]
Assessment
This is an authentic hands-on community review and stress-test of Claude Sonnet 5.5 following its launch. Bowen demonstrates live and pre-compiled software outputs across diverse languages (C++, HTML/JS, Godot/GDScript/Blender) and shows the real-time physical robot arm test without obvious deceptive edits.
Described by gemini-3.8-flash on 2026-09-29 from the video's audio and frames.