Peter Yang spent most of 2026 defaulting to ChatGPT because Claude Opus had become judgmental and started every answer with “here’s the honest truth.” He put the new Opus through five practical tests on his own channel, from 3D scene generation to editing his videos, and ends with the personality prompt that Opus 5 bombed.

Golden Gate Bridge flyover (3D from code)

  • Inspired by an OpenAI employee’s 3D model of the Palace of Fine Arts made with Astra, he pasted the same prompt and asked for a different landmark.
  • Claude drove Blender: tuning sky, tower lighting, checking keyframes frame by frame.
  • Result: a 30-second flyover, roughly 30 minutes to generate plus 30 minutes to render.
  • Same prompt to GPT Astra produced an equally impressive daytime version with cars on the bridge.
  • His takeaway: both models have essentially solved 3D scene generation.

Soaring Over the World (a playable Disney ride)

  • Prompt: “Soaring Over the World, can you show a bunch of different scenes, and let me play it in my browser.”
  • Needed one nudge - make the birds more realistic - then Claude built the 3D assets in Blender and assembled scenes: Swiss Alps, Greenland northern lights, Egyptian pyramids with hot air balloons, Fiji, the Great Wall, Paris at night with fireworks.
  • Ships as a Claude artifact you can board and fly; it also generated music for it.
  • Took about an hour, so it burns a lot of tokens.

Painting and drawing, stroke by stroke

  • “Watch Claude Paint” takes a photo and paints it like an impressionist, stroke by stroke, with styles switchable between Monet, Van Gogh and others. Output is signed “Claude.”
  • An anime counterpart draws the same photo line by line - rough circles first, then finer detail, then color.
  • A person (girl with a red umbrella on a rainy Tokyo street) came out decent, with the hair visibly duplicated.
  • Verdict: not as good as a human artist yet, but improving fast.

Computer use is still the weak spot

  • Asked to redraw his profile picture in an online MS Paint clone via the browser, the result was not impressive - and GPT Astra’s attempt made him look like a ghoul. Both bad.
  • This is the capability he most wants Anthropic to catch up on, since Codex has historically been better at browser and computer use.

Real work: design

  • Claude Design is now integrated into Claude Code: typing /design in his fitness-app project produced three layout variants in a canvas.
  • Without a design system or reference images the variants came out crowded; with specific feedback they got much better - he simplified a bloated onboarding flow to Apple/Google/email sign-in in one pass.
  • The killer part is the loop: point at a specific canvas frame (“refer to simple 1 4”) and tell Claude Code to update the app to match, with no exporting between tools.
  • Gap he flags: Claude has no image model, so swapping an AI-generated image means going back to ChatGPT - a constraint he expects to bind on design work.

Real work: video editing

  • Claude edited the intro of his Meta Muse agent video from conversation alone - animations, zoom effects, captions, and animated GIFs.
  • It also went frame by frame to blur sensitive info (his phone number, AT&T plan details) accurately - work that would take a human editor forever.
  • Tooling note: he recommends the open-source HyperFrames repo for HTML-based overlays; Claude already knows how to use it.

The personality test that Opus 5 failed

  • The prompt: review all of your chats, sources and memory, then tell me something unique you notice about me that I have not realized - it does not have to be positive, and you do not have to be nice.
  • Opus 5 produced word salad: generic, jumbled, not actionable.
  • The new model produced a specific, usable observation - his irritation is his best product strategy, and he is not applying it to his paid product - then offered concrete ideas.
  • He had a separate Claude thread compare both outputs: Opus 5 judges you, the new model explains what is going on. It still leans on a few fake-profound constructions, but far less.

For the first time in a long time, I think Claude feels like Claude again.

What he still wants: better browser and computer use in the Claude Code harness, and Anthropic thinking through new interaction patterns like the cloud-based persona agents Grok Bot has.