Claude Opus 5.5 Is INSANE – Hands-On With the BEST Model Yet!

Summary
YouTuber Bijan Bowen reviews Anthropic’s Claude Opus 5.5 release, analyzing its benchmarks, pricing structure, and safety policies before subjecting it to multiple coding and agentic benchmarks. The video evaluates Opus 5.5 across browser operating systems, full 3D games in C++ and Three.js, Godot/Blender game pipelines, a watch showcase site, and a physical robotic arm manipulation task.

What is shown

• Overview & Benchmarks [00:16 - 04:57]: Anthropic announcement page, pricing comparison ($4/$20 per million input/output tokens vs. $5/$25 on Opus 5), 1M context / 128K output specs, benchmark tables (Terminal-Bench 4.0, FrontierCode, etc.), and policy restrictions on frontier model development assistance.
• Basalt OS Test [04:58 - 11:18]: Claude Opus 5.5 (run at xHigh effort) creates a single-file browser desktop operating system featuring procedural shader wallpapers, interactive apps, a 3D voxel builder ("Voxelheim"), a 3D driving/action game ("Grand Theft Polygon"), and a multi-instance window transfer feature ("Mesh").
• C++ 3D Skateboard Game [11:48 - 14:59]: Prompted via Claude Code on Max effort, the model creates a standalone C++ NYC street skateboarding game ("Concrete Jungle") complete with trick combos, camera views, NPC collision, and pedestrian dialogue.
• Watch Brand Website [17:51 - 21:20]: At default Medium reasoning effort, Opus 5.5 builds a luxury watch showcase website with interactive 3D Three.js renders, an interactive exploded-view assembly slider, and a custom watch model textured using an uploaded photo.
• Godot & Blender 80s Wrestling Game [21:21 - 24:33]: Running on Extra effort, Opus 5.5 builds "Neon Slam '86," using Blender and Godot to generate 3D wrestler models, ring geometry, animations, crowd effects, and playable triple-threat combat mechanics.
• Robotic Arm Manipulation [24:34 - 25:53]: Opus 5.5 attempts a visual servoing task directing a robotic arm to pick up a toy truck; although it runs internal coordinate simulations, it fails to physically grasp and move the object.
• Subway Zombie FPS ("Dead Stop") [26:07 - 29:49]: A Three.js wave-based first-person shooter set in an NYC subway station featuring dynamic lighting, train arrival animations, operatic zombie vocalizations, weapon switching, and particle effects.
• Guitar Store Brawler ("Guitar Store Shred") [30:18 - 36:31]: A complex 3D simulation of a Guitar Center containing over 300 modeled instruments, playable keyboards/drums/guitars, NPC dialogue, shopper/staff anger meters, and a beat-'em-up brawl mechanic.

Claims & numbers

• The presenter says Claude Opus 5.5 performs at the level of Claude Fable 5.1 on most tasks while costing 40% less to run than Opus 5 [00:43].
• Pricing is shown as $4 per million input tokens and $20 per million output tokens for standard Opus 5.5, with cache reads at $0.20 and writes at $5 per million tokens [01:48].
• Fast mode pricing is listed at $8 per million input tokens and $40 per million output tokens [01:42].
• Model specifications show a 1,000,000 token context window and a 128,000 maximum output token limit [03:33].
• The presenter notes that Anthropic set default effort to "Medium" for Opus 5.5, while older models default to High [03:51].
• The presenter states that on Terminal-Bench 4.0, Opus 5.5 scored 66.4% compared to Fable 5.1 at 50.3%, Opus 5 at 58.0%, and GPT-6 Astra at 50.3% [01:05, 02:04].

Notable quotes

• [00:43] "it performs at the level of Claude Fable 5.1 on most work, but it costs 40% less to run than Opus 5."
• [14:44] "Can we grind on this rail? Oh. I guess not."
• [37:18] "This model is, like I think, just a game creation monster."

Assessment
This is an independent hands-on review and stress-test of Claude Opus 5.5 by an established tech creator. The demonstrations show real-time screen captures of generated code executing directly on the host machine, transparently highlighting both successes (elaborate game environments and interactive browser OS logic) and failures (inability to complete the physical robot arm grasping task).

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