As of: 2026-10-15 10:09 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/v/remakebench-opus-5-5-vs-gpt-6-astra-sonnet-sol/

# I Tested Opus 5.5 vs GPT-6 Astra, Sonnet 5.5, GPT-6.1 Sol

RemakeBench, 8 October 2026, YouTube. 144,682 views as of 15 October 2026.
Kind: Review.
Watch: https://www.youtube.com/watch?v=KW6NBXdQLDQ

## Why it is here

RemakeBench (23:42) runs four frontier models on robotics drawing, Blender scenes, Onshape CAD, PCB design, game builds and Doom; results at remakebench.com. 145k views.

## Description (written by Gemini from the video)

**Summary**
This video, created by the channel RemakeBench, presents a comparative benchmark evaluation of frontier AI models: Claude Opus 5.5, GPT-6 Astra, Claude Sonnet 5.5, and GPT-6.1 Sol. The presenter tests the models across diverse multimodal, graphics, 3D world creation, game development, decision-making, and robotic simulation tasks, comparing their generation fidelity, runtime, API cost, and efficiency.

**What is shown**
- [00:15] **Robotics Simulation Harness & Whiteboard Drawing:** Explanation of the benchmark setup in MuJoCo, followed by a physical robot arm running an Opus 5.5 controller drawing Michelangelo's *The Creation of Adam* zero-shot on a whiteboard [00:44], and a split-screen 4-way comparison drawing the Eiffel Tower [01:07].
- [01:29] **Flooded Neo-Gothic City:** 3D scene rendering in Blender comparing architectural styles and visual detail across the four models.
- [02:44] **Cathedral Interior Scene:** Comparison of interior lighting, ray propagation, and stained-glass reflections generated by each model.
- [03:44] **CAD Turbofan Design:** Step-by-step CAD assembly and cross-section modeling of a turbofan engine inside Onshape.
- [04:46] **PCB Circuit Design:** Generation of a Bluetooth speaker board in TypeScript Circuit (tscircuit), showcasing 3D PCB models, trace routing layouts, and electrical schematics.
- [05:33] **Blossom Fighting Arena (Blender):** Construction of a fighting game arena environment running a combat simulation.
- [07:05] **Three.js Space Flight Game:** Procedural generation of a space game with orbital flight, atmospheric entry, and surface terrain exploration.
- [09:31] **Snowboarding Game:** Procedural downhill slope generation with course gates, ramps, and player controls.
- [11:10] **Autonomous DOOM Gameplay:** Real-time gameplay directed by TypeSafe AI's Jev system-one decision model paired with Opus 5.5 [11:11] and GPT-6 Astra [12:33] acting as planners.
- [12:59] **Low-Poly Dioramas:** Isometric 3D Blender dioramas of a desert oasis [12:59] and a jungle temple [13:49].
- [15:29] **Robotics Manipulation Benchmarks:** Robotic manipulation in MuJoCo including simple apple pick-and-place [15:29], precision apple-stem grasping [16:36], bimanual wire puzzle disentanglement [17:08], and single/multi-stage Rubik's cube manipulation [17:36].
- [18:43] **JRPG Combat & 2D Neon Gunner Platformer:** Generation of a turn-based boss battle arena [18:43] and a 2D neon side-scrolling shooter [19:46].
- [20:29] **Sekiro Interactive Main Menu:** Stylized interactive title screens created using in-engine assets vs. 2D image-generation shortcuts.
- [21:15] **Autonomous Driving in American Truck Simulator:** Evaluation of navigation, steering, and intersection turning in *American Truck Simulator* comparing Claude Opus 5.5 [21:37] and GPT-6 Astra [22:03].

**Claims & numbers**
- **Eiffel Tower Robotic Drawing:** GPT-6 Astra finished in 4m 10s ($1.36); GPT-6.1 Sol finished in 5m 14s ($0.28); Claude Opus 5.5 finished in 7m 43s ($2.00); Claude Sonnet 5.5 finished in 28m 22s ($3.21) [01:07].
- **Flooded Neo-Gothic City:** GPT-6 Astra took 17m 6s ($4.85); GPT-6.1 Sol took 23m 55s ($0.77); Claude Opus 5.5 on High took 36m 30s ($5.46); Claude Sonnet 5.5 took 56m 43s ($14.06) [01:29]. The presenter claims using Anthropic models on "High" effort delivers virtually identical quality to "Max" while avoiding excessive token looping [02:00].
- **Cathedral Test:** Claude models took nearly two hours (Opus 5.5: 1h 42m 31s, $28.65; Sonnet 5.5: 1h 51m 32s, $27.68) compared to OpenAI models taking under 40 minutes (GPT-6 Astra: 26m 21s, $10.78; GPT-6.1 Sol: 36m 39s, $1.03) [02:44].
- **CAD Turbofan (Onshape):** GPT-6 Astra cost the most at $31.78 (57m 57s); GPT-6.1 Sol cost $3.24 (57m 17s); Claude Sonnet 5.5 cost $13.28 (44m 58s); Claude Opus 5.5 cost $14.07 (44m 59s) [03:44].
- **Blossom Fighting Arena:** Both Claude models hit usage costs exceeding $300 (Opus 5.5: $309.00; Sonnet 5.5: $312.27) before being manually halted due to high token consumption, whereas GPT-6.1 Sol finished in 3h 47m for $21.16 and GPT-6 Astra finished in 4h 28m for $178.82 [05:33].
- **Three.js Space Game:** GPT-6 Astra took 13h 38m and cost $209.16; GPT-6.1 Sol took 5h 18m costing $18.73; Claude Sonnet 5.5 took 5h 7m costing $46.41; Claude Opus 5.5 took 2h 10m costing $47.50 [07:05].
- **Jev DOOM Decision Model:** The presenter states Jev operates with a measured latency of around 190 ms per call [11:31]. During the test runs, Opus 5.5 achieved 25 kills and 4 deaths [12:41], while Astra achieved 24 kills and 5 deaths [12:45].
- **Bimanual Wire Puzzle:** Both Claude models failed due to physics engine collision clipping, whereas both OpenAI models passed [17:18].
- **American Truck Simulator:** Opus 5.5 failed after missing its designated turn [21:54]; GPT-6 Astra navigated through depot gates and turned onto the public road, but caused a major collision at an intersection by crossing into opposing traffic [22:45].

**Notable quotes**
- [01:00] *"And, yeah, again, this is zero-shot, you know, there's no training, there's no policy, there's no policy training, so it's all zero-shot with an LLM."*
- [02:00] *"What I found is on Max, it's a bit overkill. The model will just keep going and going and going and going and won't really stop..."*
- [20:39] *"The GPT models used image-gen and built effects around the image that they generated, whereas the Claude models, given they don't have image-gen, they built the whole scene from scratch."*

**Assessment**
This is an authentic, independent benchmark review video systematically evaluating Claude and GPT models across practical robotics, coding, and 3D simulation workflows. The demonstrations present real code execution, simulator telemetry, terminal logs, and live gameplay capture, objectively documenting both successes and severe failure modes across all evaluated models.

_Described by gemini-3.8-flash on 2026-10-15 from the video's audio and frames._

## Related

- 2026-10-08: [OpenAI ships Ultrafast mode for GPT-6.1 Sol in the API, Codex and ChatGPT Work](https://postcutoff.com/e/2026-10-08-gpt-6-1-sol-ultrafast/)
- 2026-09-22: [Anthropic releases Claude Opus 5.5](https://postcutoff.com/e/2026-09-22-claude-opus-5-5/)
- 2026-09-03: [OpenAI releases GPT-6 Astra, its first GPT-6 model](https://postcutoff.com/e/2026-09-03-gpt-6-astra/)