The Build-Off: Three Prompts, Four Models, One Winner per Round
TryAI ran a controlled test: give each model the exact same three prompts, one shot each (with one allowed retry if the app didn't render), and evaluate the raw output in a real browser. The models: Grok 4.5, GPT-5.5, Claude Opus 4.8, and Claude Fable 5.
Round 1: 3D Rubik's Cube with Scramble and Solve
This tests 3D math, per-face state management, and animation — all in a single HTML file. The Claudes won outright: both Opus 4.8 and Fable 5 produced a correctly colored, animated cube on the first try. Grok 4.5's first attempt rendered only the title and buttons (blank void), but the retry produced a clean cube. GPT-5.5 failed: it rendered a single dark face with barely any color. Winner: Claude Opus 4.8 and Fable 5 (tie).
Round 2: Particle Gravity Sandbox
All four shipped a working sandbox. The decision came down to visual quality. GPT-5.5's version was the most mesmerizing: glowing neon attractors with dense, swirling colored trails. Grok 4.5's was clean and orbital with tidy attractor rings. Fable 5 used soft glowing orbs. Opus 4.8 had great physics but less eye-candy. Winner: GPT-5.5 (on vibes).
Round 3: Playable Breakout Game
All four produced a polished, playable brick-breaker with score, lives, and a glowing paddle on the first try. Grok 4.5 and GPT-5.5 leaned into neon arcade aesthetics. Winner: everybody.
Speed and Cost Benchmarks
TryAI measured latency, throughput, and cost using three fixed prompts (coding, reasoning, summarization), three reps each, capped at 400 output tokens. Results:
| Model | Median Latency | First Token | Throughput | Cost / Reply | Success |
|---|---|---|---|---|---|
| Grok 4.5 | 2.8s | 0.44s | 110 tok/s | $0.002 | 100% |
| GPT-5.5 | 2.0s | 1.26s | 53 tok/s | $0.004 | 100% |
| Claude Opus 4.8 | 2.6s | 1.16s | 47 tok/s | $0.004 | 100% |
| Claude Fable 5 | 6.3s | 3.47s | 28 tok/s | $0.009 | 100% |
Grok 4.5 hit first token in under half a second (0.44s) and streamed at ~110 tokens/second — roughly double the throughput of any other model. It was also the cheapest at $0.002 per reply. However, its median latency was mid-pack because it wrote the most tokens per answer, and its p95 latency spiked to ~9s. GPT-5.5 was snappiest on short answers, Opus 4.8 sat in the middle, and Fable 5 was the slowest and priciest.
Bonus Round: SVG of a Horse Riding an Astronaut on the Moon
Fable 5 stole the show with a cowboy-hatted horse yelling "Giddy-up, human!" while a hunched astronaut wheezes. GPT-5.5 was close second with a gleeful horse shouting "WHEE!". Grok 4.5 produced a clean, readable scene. Opus 4.8 shipped a duplicate attribute in raw SVG that trips strict parsers (rendered leniently in the test).
The Verdict
Grok 4.5 is the speed-and-value monster: great for high-volume codegen where latency and cost compound. Its only miss was the hardest stateful task (blank cube on first try, fixed on retry). Claude Opus 4.8 and Fable 5 are the most reliable builders, nailing the 3D cube first try — but you pay in latency and price. GPT-5.5 is the snappy stylist with prettiest gravity sim, but flubbed the cube.
Next step: Run your own prompts on TryAI (pay-as-you-go, no subscription) to compare these models on your specific workload.
