Skip to content

The new gpt-6.1 terra beat sonnet and kimi – 3.5x cheaper and 3.5x faster than kimi

Illustrated Hachiko dog statue in a witch hat next to a glowing jack-o'-lantern on a neon Halloween night in Shibuya models

The new GPT-6.1 Terra built Shibuya on Halloween night in three.js for $0.95 in 31 minutes. Claude Opus 5.5, Claude Sonnet 5.5 and Kimi K3.1 Pro got the same brief. Terra was the cheapest and the fastest of the four, while Opus 5.5 had the most cinematic light.

GPT-6.1 Terra, Claude Opus 5.5, Kimi K3.1 Pro and Claude Sonnet 5.5 build Shibuya on Halloween night in three.js

The setup

Every model got the same task, written as a chunked brief, and went through the same stages:

  1. The main scene – Shibuya at blue hour: the Scramble Crossing, the giant screens, the Hachikō statue, Center-Gai and the view over Tokyo. Each model wrote a plan, the code in chunks, then three rounds of comparing its renders with five reference photos
  2. Halloween – a costumed crowd: witches, vampires, skeletons, ghosts, pumpkin heads, mummies, with pumpkin lanterns and garlands
  3. A cosplay close-up and a breakdancer – a Gojo Satoru cosplayer in a close-up at the start, and a breakdancer on Center-Gai

All four worked from the same text, and none saw another model's scene.

  • Models: GPT-6.1 Terra, Claude Opus 5.5, Claude Sonnet 5.5, Kimi K3.1 Pro
  • Terra and Kimi through OpenRouter, Opus and Sonnet through a Claude subscription
  • One scene in plain HTML and three.js, everything generated in code
  • Five camera shots, 7 seconds each, recorded the same way for all four
  • Test date: Oct 26, 2026

The reference photos came from Wikimedia Commons and were only shown to the models, not used in the scenes.

The numbers

Terra's and Kimi's costs are what OpenRouter billed. Opus and Sonnet ran on a subscription, so their costs are estimates at the API list price.

ModelCostTime
GPT-6.1 Terra$0.9531m
Claude Sonnet 5.5~$2.301h 00m
Kimi K3.1 Pro$3.281h 50m
Claude Opus 5.5~$4.901h 12m

Terra was about 5x cheaper than Opus by our estimate and 3.5x faster than Kimi. OpenRouter lists it at $1 per million input tokens and $5 per million output. Our full review of the model is in GPT-6.1 Terra: half the price of Sol.

Five shots, four Shibuyas

Each frame below shows Terra and Opus on top, Kimi and Sonnet below.

Four 3D Shibuya scenes at night, each opening on a white-haired blindfolded cosplayer in a Halloween crowd
The opening close-up: the cosplayer in all four scenes
The Scramble Crossing seen from above in four 3D scenes, with zebra stripes, neon screens and a costumed crowd
The Scramble Crossing from above
Four 3D versions of the QFRONT building and the giant video screens over the Shibuya crossing at blue hour
The giant screens
The Hachiko dog statue in four 3D scenes, surrounded by people in Halloween costumes
Hachikō and the Halloween crowd
The view over Tokyo from Shibuya Sky in four 3D scenes, from a grey city block grid to a pink hazy skyline
The view over Tokyo
  • GPT-6.1 Terra – the cleanest crossing and readable screens, but a flat grey city in the final view and simple blocky people
  • Claude Opus 5.5 – the most cinematic light: wet reflections, glowing screens and light trails across the city at night
  • Claude Sonnet 5.5 – the brightest neon and the most colorful crowd, with a soft-focus Hachikō and a pink haze over Tokyo
  • Kimi K3.1 Pro – a dense crowd at Hachikō and a sharp crossing, with the plainest buildings and the same grey city view as Terra

The most expensive scene looks the best, and the cheapest one is not far behind on the crossing and the screens.

Our take

GPT-6.1 Terra is the new price and speed leader for this kind of job. It built a full Shibuya with a Halloween crowd for under a dollar, in half an hour. Opus 5.5 still renders the night most like a film, at about five times the price. If you need many scenes fast, Terra is the one to try first.

Sources

Portrait of Yana Lialuke, author at thehype.news

Yana Lialuke

Content Creator at thehype. Yana tests AI models hands-on across image generation, 3D, and LLMs. Same prompts, real costs. An AI marketer turned creator, she builds with agents, from 3D games to Three.js scenes, sharing prompts and failures.

More articles by Yana Lialuke →
ON AIR · RADIO.THEHYPE.NEWS ↗ ai news radio — 24/7