Opus 5.5 vs GPT-6 Astra build the same physics Rube Goldberg machine: cost and time compared

Summary
This side-by-side comparison video, posted by user @leogao25 on September 22, 2026, compares Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra coding and executing a self-running 3D physics Rube Goldberg machine from scratch. Both models were given a single high-effort run from an empty directory using their respective developer harnesses (Claude Code vs. Codex CLI), with the screen showing the resulting 3D simulations alongside performance, token, and cost metrics.

---

What is shown

• Prompt & benchmark setup: Text at the top states: “Opus 5.5 vs GPT-6 Astra: same prompt, same machine” with the prompt: “Build a self-running Rube Goldberg machine. Physics from scratch. One attempt per model, high effort.” [00:00].
• Left pane (Claude Opus 5.5 via Claude Code):
  • [00:00–00:06]: A metal sphere drops down a multi-level wooden ramp within a detailed wooden room/cabinet enclosure.
  • [00:07–00:10]: The ball rolls across a wooden tabletop, striking and toppling a line of dominoes.
  • [00:11–00:16]: The falling domino triggers a counterweight pulley system that raises a checkered finish flag up a flagpole.
• Right pane (GPT-6 Astra via Codex CLI):
  • [00:00–00:06]: A marble rolls through elevated green curved wireframe tracks on a minimalist green plane.
  • [00:07–00:10]: The marble trips a row of standing upright pegs and a seesaw rocker lever.
  • [00:11–00:16]: The kinetic impulse triggers a pulley contraption that hoists an orange triangular target flag.
• Comparative performance table: Detailed breakdown showing build time, API cost, output token count, and pricing tiers [00:00–00:16].

---

Claims & numbers

• Claude Opus 5.5:
  • Harness: Claude Code.
  • Build time: 62 minutes.
  • API cost: $14.28.
  • Output tokens: 265K.
  • List price per million tokens: $4 input / $20 output.
• GPT-6 Astra:
  • Harness: Codex CLI.
  • Build time: 11 minutes.
  • API cost: $2.20.
  • Output tokens: 20K.
  • List price per million tokens: $10 input / $50 output.
• Ratio (Opus 5.5 relative to Astra):
  • Build time: 5.8× longer.
  • API cost: 6.5× higher.
  • Output tokens: 13.6× more.
  • List price ratio: 0.4× (cheaper list price per token).
• Run conditions: Stated at the bottom: “Each model started in an empty folder with no config files. Cost is API list price for every token used, including the agent’s pre-training. Time is wall-clock from prompt to finished file. One run each.”

---

Notable quotes

• “Build a self-running Rube Goldberg machine. Physics from scratch. One attempt per model, high effort.” [00:00]
• “Each model started in an empty folder with no config files.” [00:00]
• “Time is wall-clock from prompt to finished file. One run each.” [00:00]

---

Assessment
This is a real head-to-head developer benchmark of two frontier models building complex interactive simulations autonomously. While both successfully produced functioning physics contraptions from a cold start, the test reflects a single trial (One run each) and compares two distinct agent harnesses (Claude Code vs. Codex CLI), which significantly influences token volume, reasoning loops, and wall-clock time alongside raw model capability.

---

Lyrics & themes

• The video is completely silent with no audio, narration, or soundtrack.
• Key themes focus on real-world coding agent cost efficiency, speed versus aesthetic depth, and autonomous physics engine scripting.

---

Lore & references

• The “Price Reversal” Phenomenon: Illustrates a known dynamic where a model with a significantly lower token list price ($4/$20 for Opus 5.5 vs. $10/$50 for Astra) ends up being drastically more expensive overall ($14.28 vs. $2.20) due to deeper reflection, multi-step debugging, and verbose output token generation.
• Claude Code vs. Codex CLI: Highlights the ongoing ecosystem rivalry between Anthropic’s terminal agent and OpenAI’s terminal CLI harness.

---

Visual style & craft

• Layout: A clean, structured side-by-side benchmark dashboard with synchronized 3D viewport recordings above a data comparison table.
• Visual generation: The simulations are rendered in 3D WebGL/canvas scenes coded by the models. Opus 5.5 prioritized richer environment modeling (textured wooden walls, lighting, polished surfaces), while GPT-6 Astra wrote a streamlined, minimalist, functional wireframe structure that solved the objective with fewer tokens and in a fraction of the time.

Described by gemini-3.8-flash on 2026-10-10 from the video’s audio and frames.