Peter Yang compares Gemini 4 Neon, Claude Opus 5.5 and Gemini Carbon across five familiar tasks in this 18-minute walkthrough.

Which Gemini model he would use

  • Yang picks Neon for most coding, writing and automation, reserving Carbon for unusually demanding projects and Hydrogen for quick questions.
  • That is his working preference. The demonstrations below show specific outputs and trade-offs.

1. Shibuya Crossing 3D video

  • Neon and Opus each took about 40 minutes in his runs; Carbon took 73 minutes.
  • Yang prefers Opus’s dramatic zoom, train and busier crossing, despite visibly artificial people.
  • Neon’s reflections look good, but its crossing feels small and some walkers have static legs.
  • Carbon’s scene has less convincing street geometry and fewer of the details he wanted.

2. Godzilla game

  • Neon runs smoothly, with simple graphics and a breath effect that appears to emerge from the creature’s neck.
  • Opus adds a more atmospheric introduction, richer destruction and more effects, but plays noticeably slower on Yang’s MacBook M5.
  • Carbon’s camera behavior prevents him from properly seeing the character move.
  • Yang gives Opus the visual win while acknowledging Neon’s better responsiveness.

3. Website redesign

  • Neon produces a clearer landing page than Carbon, with a call to action and an animated visual.
  • Carbon overloads the page with text and decorative elements.
  • Yang still prefers Opus’s bolder motion design, clothes-hanger interface and drawers.

4. Writing and revision

  • All three write a fictional software engineer’s reflection on no longer typing code manually.
  • Yang narrowly ranks the essays Neon, Carbon, then Opus, while calling the results close and taste-dependent.
  • The reveal is only partly blind: he admits recognizing Carbon’s essay from an earlier video.
  • Neon also gives concise revision bullets and before-and-after examples; Carbon’s feedback is harder for him to scan.

5. Advice under pushback

  • Neon recommends a separate interview channel; Carbon proposes dropping interviews; Opus recommends timely solo tests plus selective guests.
  • Yang prefers Opus’s advice, while acknowledging that he personally enjoys making interviews.
  • He reports pushing back five ways and finding broadly similar resistance across models, with real data prompting revisions.
  • The walkthrough shows selected exchanges, so it supports a small qualitative comparison rather than a measured resistance score.

Keep the workflow portable

  • Yang’s broader recommendation is to retain skills and workflows in a personal repository that works across models and agent software.
  • His conclusion is narrower than a universal ranking: Neon feels competitive for everyday writing and advice; Opus remains his visual-design favorite.
  • Caveat: these are creator-presented examples and personal judgments, with no independently reproduced benchmark or repeated-run statistics.

“The most durable thing that you can build is your own personal system of AI skills and workflows.” — Peter Yang, 16:31